Low power factor load scene classification method based on KAN network

Abstract

The invention discloses a low-power-factor load scene classification method based on a KAN network, and belongs to the field of load scene classification, and the method comprises the steps: collecting a data feature set of an electric energy meter; mining hidden correlation coefficients among features in the electric energy meter data feature set; carrying out feature intersection on the two feature quantities of which the hidden correlation coefficients are greater than a hidden correlation threshold to obtain new feature data; training a CNNKAN network load scene classification model by using the new feature data to obtain a CNNKAN network load scene classification model meeting performance requirements; and deploying the CNNKAN network load scene classification model meeting the performance requirements into electric energy meter software, and carrying out load scene classification on electric energy meter data of a scene to be tested. According to the method, the problem that the classification result is not accurate enough due to the fact that interaction and overall influence among multiple factors are ignored in an existing method is solved.

Description

Technical Field

[0001] The present invention belongs to the field of load scenario classification, and in particular relates to a low power factor load scenario classification method based on a KAN network. Background Art

[0002] As a core technology system ensuring fair energy trading and the economic operation of power grids, the accuracy of electricity metering and traceability directly impacts the protection of social public welfare rights, the optimization of power grid efficiency, and the industrialization of smart grids. As a legally standardized metering device used by both power suppliers and users, the performance of electricity meters remains a key technical indicator for the operation of the power market. However, with the transformation and upgrading of the power system, electricity metering faces new challenges.

[0003] The traditional power system, primarily based on conventional power generation, is transitioning to a new power system centered on renewable energy. This shift has fundamentally altered the grid's operational structure and the load characteristics on both the generation and demand sides, resulting in a system characterized by "double highs"—high penetration of renewable energy and high penetration of power electronics. The dynamic reconfiguration of the system topology leads to significant differences in load characteristics between the generation and consumption sides. On the generation side, distributed generation resources (such as photovoltaic power plants and wind farms) exhibit significant intermittency due to meteorological constraints. On the demand side, the large-scale integration of power electronic loads (such as high-speed rail traction systems and electric vehicle charging station clusters) generates a wide range of dynamic loads due to fluctuating operating conditions. This structural change leads to extreme variations in the system's power factor (PF). For example, the PF of high-speed rail systems and charging stations often drops to 0.1 under light load or transient conditions. However, existing electricity meters do not account for the PF range of 0.01-0.25, resulting in inaccurate measurements at low PF levels. Furthermore, fundamental research on load characteristics and modeling in different scenarios under low PF conditions remains insufficient. Therefore, it is of great theoretical and practical significance to classify and study low PF scenarios and clarify their load characteristics.

[0004] Load scenario classification is a fundamental and critical technology for the refined management and optimized dispatch of smart grids. Similar to the challenges faced in smart meter error assessment, load scenario classification requires considering the combined effects of multiple factors, including but not limited to the type and operating status of electrical equipment, environmental factors (such as temperature and humidity), temporal factors (such as season and time of day), and user behavior patterns. These factors intertwine and jointly influence the load characteristics of the power grid, complicating tasks such as load forecasting, demand-side management, and the optimal allocation of distributed energy resources. However, existing load scenario classification methods often focus on analyzing a single or a few factors, ignoring the interactions and overall impact of multiple factors. This results in inaccurate classification results and is unable to meet the requirements of refined smart grid management. Furthermore, traditional classification models such as decision trees and support vector machines have limitations when processing high-dimensional and nonlinear data, making it difficult to effectively capture the complex characteristics of load data.

[0005] Existing classification methods have the following shortcomings: 1. Complexity of data acquisition and processing.

[0006] Uneven data quality: Due to the diverse data sources, there may be inconsistencies in data quality, such as missing data, outliers, noise, etc., which may have a negative impact on the accuracy of load scenario classification.

[0007] Difficulty in data processing: Load data often has complex characteristics such as high dimensionality and nonlinearity, making it difficult to process. A series of preprocessing tasks, such as data cleaning, feature extraction, and dimensionality reduction, are required to improve the performance and efficiency of the classification model.

[0008] 2. Limitations of modeling methods and model selection.

[0009] Single modeling approach: Many current load scenario classification technologies still rely on traditional statistical methods or machine learning algorithms, which can exhibit limitations when processing complex and variable load data. These methods fail to specify specific load modeling methods or load models for different load scenarios or types, instead employing a single, universal approach and model.

[0010] Poor model adaptability: Load scenario classification technology needs to be highly adaptable to cope with the changing characteristics of different power systems and loads. However, existing models may require extensive adjustments and optimizations when dealing with new loads and application scenarios, and may even be incapable of direct application.

[0011] 3. Real-time and computing efficiency issues.

[0012] High real-time requirements: In power systems, load forecasting and classification often require high real-time performance to enable timely adjustments to power dispatch and distribution strategies. However, existing technologies, when processing large amounts of load data, can suffer from high computational complexity and slow processing speeds, making it difficult to meet real-time requirements.

[0013] Limited computing resources: In practical applications, computing resources are often limited. How to achieve efficient and accurate load scenario classification within limited computing resources is a major challenge facing current technologies. Summary of the Invention

[0014] In view of the above-mentioned deficiencies in the prior art, the present invention provides a low power factor load scenario classification method based on a KAN network, which solves the problem that the existing method ignores the interaction and overall impact between multiple factors, resulting in inaccurate classification results.

[0015] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a low power factor load scene classification method based on the KAN network, comprising: Collect electricity meter data and time-align the data of each electricity meter to obtain a sample data set; The data of each electric energy meter in the sample data set is processed in sequence by removing abnormal data, filling in missing data and normalizing data to obtain a preprocessed data set; Perform feature extraction on each data in the preprocessed data set to obtain several groups of feature value samples, and integrate the feature value samples to obtain the electric energy meter data feature set; Mining the hidden correlation coefficients between the features in the electric energy meter data feature set; Perform feature crossover on two feature quantities whose hidden correlation coefficient is greater than the hidden correlation threshold to obtain new feature data; Using the new feature data, the CNN_KAN network load scenario classification model is trained to obtain a CNN_KAN network load scenario classification model that meets the performance requirements; The CNN_KAN network load scenario classification model that meets the performance requirements is deployed in the electricity meter software, and the load scenario classification is performed on the electricity meter data of the test scenario.

[0016] Furthermore, the electric energy meter data includes voltage data, current data, active power data, power factor data and scene position data.

[0017] Furthermore, the abnormal data elimination process adopts the box method; the missing data filling process adopts the polynomial function method.

[0018] Furthermore, the characteristic value samples include time domain features and frequency domain features; the time domain features include mean value, standard deviation, skewness, kurtosis, maximum value, minimum value, peak-to-peak value, root mean square, rate of change, amplitude factor, waveform factor and impulse factor; the frequency domain features include rate spectrum density and spectrum centroid.

[0019] Furthermore, the hidden correlation coefficient between the features is calculated as follows: Based on the feature set of the electric energy meter data, the correlation coefficient between each feature and the load is calculated:

[0020] in, For load and features The correlation coefficient between Characterized by No. eigenvalues; Characterized by The average eigenvalue of For the The load value corresponding to the eigenvalue sample; is the average load value; is the total number of eigenvalue samples; is the eigenvalue sample index; According to the correlation coefficient between each feature and the load, the hidden correlation coefficient between the features is calculated:

[0021] in, Characterized by and features The hidden correlation coefficient between For load and features The correlation coefficient between Characterized by and features The correlation coefficient between For load and features The correlation coefficient between .

[0022] Furthermore, the new feature data includes several new feature quantities:

[0023] in, For the The new feature quantity is calculated by the eigenvalue samples; Characterized by No. dimensional latent vector; is the feature index; For the The first eigenvalue sample that participates in feature crossover eigenvalues.

[0024] Furthermore, the CNN_KAN network load scenario classification model is specifically a CNN convolutional neural network; the fully connected layer of the CNN convolutional neural network is replaced by a KAN network.

[0025] The beneficial effects of the present invention are as follows: first, based on the actual measured data of the electricity meters, the present invention mines the deep connections between the data of each electricity meter through feature cross-mining, thereby constructing a new feature quantity. Utilizing the excellent automatic feature extraction capability of CNN, rich spatial features are extracted from the load data, aiming to enhance the stability of the model and improve the training efficiency. Subsequently, the unique advantages of KAN in capturing key information or processing specific tasks are utilized to further explore the key features in the data. Compared with traditional methods, the CNN_KAN network model proposed in the present invention shows higher accuracy and better stability in the load scenario classification task under low power factor. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 Flow chart of the method of the present invention.

[0027] Figure 2 FIG. 4 is a network structure diagram of KAN in this embodiment. DETAILED DESCRIPTION

[0028] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0029] like Figure 1 As shown, in one embodiment of the present invention, a low power factor load scenario classification method based on a KAN network includes: Collect electricity meter data and time-align the data of each electricity meter to obtain a sample data set; The data of each electric energy meter in the sample data set is processed in sequence by removing abnormal data, filling in missing data and normalizing data to obtain a preprocessed data set; Perform feature extraction on each data in the preprocessed data set to obtain several groups of feature value samples, and integrate the feature value samples to obtain the electric energy meter data feature set; Mining the hidden correlation coefficients between the features in the electric energy meter data feature set; Perform feature crossover on two feature quantities whose hidden correlation coefficient is greater than the hidden correlation threshold to obtain new feature data; Using the new feature data, the CNN_KAN network load scenario classification model is trained to obtain a CNN_KAN network load scenario classification model that meets the performance requirements; The CNN_KAN network load scenario classification model that meets the performance requirements is deployed in the electricity meter software, and the load scenario classification is performed on the electricity meter data of the test scenario.

[0030] The purpose of this invention is to establish a load scenario classification model using a CNN-KAN network. This method first uses measured electricity meter data to explore deep connections between individual meter data through feature cross-pollination, thereby constructing new feature quantities. Secondly, the convolutional neural network (CNN) is improved and integrated with the KAN network. This overcomes the CNN model's focus on local features, as well as the gradient vanishing or gradient exploding problems and increased overfitting risk associated with increasing network depth. By establishing a load scenario classification model using a CNN-KAN network, it is possible to accurately determine load scenarios and assess current operating status.

[0031] Collect electricity meter data. This includes the following specific steps: Collect and align data. Obtain sample data from the energy meter; the sample data includes voltage data, current data, active power data, power factor data, and scene location data, and time-align data collected at different frequencies.

[0032] The electric energy meter data includes voltage data, current data, active power data, power factor data and scene position data.

[0033] The abnormal data elimination process adopts the box method; the missing data filling process adopts the polynomial function method.

[0034] The characteristic value samples include time domain features and frequency domain features; the time domain features include mean value, standard deviation, skewness, kurtosis, maximum value, minimum value, peak-to-peak value, root mean square, rate of change, amplitude factor, waveform factor and impulse factor; the frequency domain features include rate spectrum density and spectrum centroid.

[0035] In this embodiment, data preprocessing includes the following specific steps: A box plot is used to identify outliers in the electricity meter data. The specific criteria for identification are to calculate the minimum and maximum estimated values ​​in the data. If the data exceeds this range, it indicates that the value is likely an outlier. Data identified as outliers are then treated as missing values.

[0036] Energy meter data often contains missing values ​​due to various reasons, including irregular meter management, irregular data collection processes, and irregular data recording. Based on the characteristics of energy meter data, this paper uses a polynomial function to fill in missing values ​​(polynomial interpolation algorithm). This method substitutes existing raw data points into a formula to calculate the unknown missing values.

[0037] Data normalization: For the i-th historical measurement value at the k-th time step, perform data normalization on the historical measurement value.

[0038] In this embodiment, the extracted electric energy meter data features include: time domain features and frequency domain features.

[0039] Time domain features: Time domain features are extracted based on the performance of the signal in the time domain. These features usually include the statistical properties of the signal, such as mean, variance, standard deviation, skewness, kurtosis, etc., which describe the basic statistical information of the signal. In addition, time domain features may also include the waveform characteristics of the signal, such as peaks, troughs, zero crossings, etc., as well as the changing trend and periodicity of the signal. Time domain features can intuitively reflect the temporal changes and dynamic characteristics of the signal. The time domain features used in the present invention are mean, standard deviation, skewness, kurtosis, maximum value, minimum value, peak-to-peak value, root mean square, rate of change, amplitude factor, waveform factor and impulse factor, which are defined as follows: Mean: The average value of a signal in the time domain, used to describe the central trend of the signal.

[0040] Standard Deviation: The degree of dispersion of signal values, used to measure the volatility of the signal.

[0041] Skewness: The asymmetry of the signal distribution, used to describe the skew direction of the signal distribution.

[0042] Kurtosis: The sharpness of the signal distribution, used to describe the peak state characteristics of the signal distribution.

[0043] Maximum: The maximum value of a signal in the time domain, used to reflect the peak strength of the signal.

[0044] Minimum: The minimum value of the signal in the time domain, used to reflect the lowest signal strength.

[0045] Peak-to-Peak: The difference between the maximum and minimum values ​​of a signal, used to reflect the overall fluctuation range of the signal.

[0046] Root Mean Square (RMS): The effective value of a signal, which reflects the energy level of the signal.

[0047] Rate of Change: The rate of change of a signal over time, used to reflect the dynamic change speed of the signal.

[0048] Amplitude Factor: The ratio of the signal peak value to the effective value, which is used to reflect the relationship between the peak value and energy of the signal.

[0049] Form Factor: The ratio of the effective value of a signal to its average value, used to reflect the waveform characteristics of the signal.

[0050] Crest Factor: The ratio of the signal peak value to the effective value, which is used to reflect the relationship between the peak value and energy of the signal, and is particularly suitable for detecting impact signals.

[0051] Frequency domain features: Frequency domain features are extracted based on the performance of the signal in the frequency domain. These features are usually obtained through spectral analysis of the signal, such as power spectral density, spectral centroid, spectral spread, etc. Frequency domain features describe the energy distribution and spectral characteristics of the signal at different frequencies. Through frequency domain features, we can understand information such as the frequency component, bandwidth, harmonic structure, and symmetry of the spectrum of the signal. The frequency domain features used in this invention are power spectral density and spectral centroid, and their definitions and formulas are as follows: Power Spectral Density: The distribution of signal power in the frequency domain, indicating the energy intensity of the signal at different frequencies.

[0052] Spectral Centroid: An indicator of the center position of the signal spectrum energy distribution.

[0053] The calculation of the hidden correlation coefficient between the features is specifically as follows: Based on the feature set of the electric energy meter data, the correlation coefficient between each feature and the load is calculated:

[0054] in, For load and features The correlation coefficient between Characterized by No. eigenvalues; Characterized by The average eigenvalue of For the The load value corresponding to the eigenvalue sample; is the average load value; is the total number of eigenvalue samples; is the eigenvalue sample index; According to the correlation coefficient between each feature and the load, the hidden correlation coefficient between the features is calculated:

[0055] in, Characterized by and features The hidden correlation coefficient between For load and features The correlation coefficient between Characterized by and features The correlation coefficient between For load and features The correlation coefficient between .

[0056] The new feature data includes several new feature quantities:

[0057] in, For the The new feature quantity is calculated by the eigenvalue samples; Characterized by No. dimensional latent vector; is the feature index; For the The first eigenvalue sample that participates in feature crossover eigenvalues.

[0058] In this embodiment, the extracted feature values ​​undergo correlation analysis, including internal analysis of time-domain features, internal analysis of frequency-domain features, and cross-domain analysis. Hidden correlation coefficients are calculated to determine whether there is a deep correlation between the feature and the predicted value. Feature values ​​with deep connections are then cross-referenced to obtain new feature values.

[0059] The CNN_KAN network load scenario classification model is specifically a CNN convolutional neural network; the fully connected layer of the CNN convolutional neural network is replaced by a KAN network.

[0060] In this embodiment, convolutional neural networks (CNNs), one of the core algorithms in deep learning, play a crucial role in image processing and computer vision tasks due to their powerful feature extraction capabilities. The CNN network structure consists of three main components: an input layer, hidden layers, and an output layer. The hidden layers are further subdivided into convolutional layers, pooling layers, and fully connected layers. This hierarchical design enables CNNs to efficiently and accurately extract and integrate key features from input data.

[0061] The input layer directly receives raw data and is the starting point for data to enter CNN for processing.

[0062] The convolutional layer is the core of a CNN. It slides a defined convolution kernel over the input data, performing local connections and weight sharing. This process aims to extract local features of the input data, such as edges and textures. A nonlinear activation function is typically used after the convolutional layer to enhance the network's nonlinear representation capabilities, enabling it to learn more complex features. In this paper, the ReLU function is used as the activation function.

[0063] The pooling layer follows the convolutional layer. Its primary purpose is to reduce the dimensionality of the feature map, essentially reducing its height and width while retaining important information. By compressing redundant information in the feature map, pooling simplifies the network's computational complexity, accelerates training, and, to a certain extent, prevents overfitting. Common pooling methods include max pooling and average pooling. This application uses max pooling.

[0064] The fully connected layer is located at the end of the CNN and is used to integrate the local features extracted by the previous convolutional layer and pooling layer to form a global overall feature representation. In the fully connected layer, each node is connected to all the nodes in the previous layer. This fully connected method enables the network to learn the complex relationship between the input data and output the final prediction result. However, it should be noted that in some modern CNN architectures, in order to reduce the number of parameters and prevent overfitting, the fully connected layer may be replaced by a global average pooling layer or other more efficient layers. In this invention, the KAN network is used to replace the fully connected layer of the original CNN network.

[0065] Kolmogorov-Arnold Networks (KANs), an emerging neural network architecture, have emerged in the fields of complex function approximation and data processing with their unique theoretical background and practical potential. KANs are inspired by the Kolmogorov-Arnold representation theorem, which mathematically states that any multivariable function can be expressed as a series of single-variable functions combined with addition and compound operations. This theoretical framework provides a solid mathematical foundation for the construction of KANs.

[0066] The network structure of KAN also follows the hierarchical design principle, but compared with the traditional convolutional neural network, it has a more unique hierarchical configuration. The core of KAN includes the input layer, the composite layer, the single variable function layer and the output layer. These layers work together to achieve a complex mapping from input data to output results. Its network structure is as follows Figure 2 shown.

[0067] The input layer receives the data to be processed. This data can be multidimensional and contain various features and variables. As the starting point of the KAN processing flow, the input layer ensures that the data can enter the network seamlessly.

[0068] The composite layer is a key component of the KAN, responsible for performing composite operations on data. Based on the Kolmogorov-Arnold representation theorem, the composite layer simulates the interactions between different variables through specific connections. These interactions affect the final output in a nonlinear manner. The design of the composite layer enables the KAN to capture the complex interdependencies in the input data.

[0069] The univariate function layer follows the composite layer and is responsible for applying univariate functions to the data after the composite operation. These univariate functions can be any differentiable functions, such as Sigmoid, Tanh, or ReLU. They enhance the nonlinear modeling capabilities of the KAN, enabling the network to learn more complex data patterns.

[0070] The output layer is the terminal of KAN, integrating the features extracted and transformed by the previous layers to generate the final output result. Depending on the specific task, the output layer can take different forms, such as softmax output for classification tasks, linear output for regression tasks, etc. In this application, the softmax output of the classification task is used. The obtained new feature quantity is input into the improved load scenario classification model for training. The specific process is as follows: The input one-dimensional data of length 12 is passed through three convolutional layers to further deepen feature learning and extract more abstract and representative features. The ReLU activation function is also applied to the output of each hidden layer. This ReLU activation function in the hidden layer further enhances the nonlinear characteristics of the network, enabling the model to better fit complex data distributions. The features are then downsampled through a pooling layer to reduce the data dimension and computational complexity while retaining the most important feature information. The feature map output by the pooling layer is flattened into a one-dimensional vector and input to the next network layer. Finally, after processing through three convolutional and pooling layers, the CNN network output is input into the KAN network. The softmax activation function converts the neuron output into a probability distribution, thereby achieving load scenario classification and outputting the load scenario category.

[0071] During the model training phase, the present invention uses the AdamW optimizer. The AdamW optimizer is popular for its adaptive learning rate feature, which can dynamically adjust the learning rate based on the model's training status, thereby improving the model's convergence speed and performance. A training period of 200 epochs was set to ensure that the model had sufficient time to learn the characteristics of the data. During the backpropagation process, the cross-entropy loss function was used. The cross-entropy loss function is a common method for measuring the difference between model predictions and true labels in classification problems. By minimizing the cross-entropy loss, the probability distribution of the model predictions can be made closer to the distribution of the true labels, thereby improving classification accuracy.

[0072] Model performance evaluation. The data collected in the test set is input into the trained load scenario classification model to obtain load classification results. Based on the load classification results and the actual load categories, the model's accuracy, precision, recall, and F1 value are calculated. These are used as indicators to evaluate model performance. The load scenario classification model is considered to have passed the test when all four indicators meet the threshold. If the test and evaluation results meet the requirements, the trained load scenario classification model is deployed. Otherwise, the load scenario classification model is adjusted and optimized.

[0073] Deploy the load scenario classification model that has passed training and testing: Deploy the load scenario classification model that has passed training and testing to the electricity meter software; Real-time collection of voltage data, current data, active power data, power factor data and scene location data of the scene to be tested; Input the real-time collected test scenario data into the load scenario classification model; The load scenario classification model outputs corresponding prediction results. While outputting the corresponding prediction results, real-time monitoring is achieved through the trained and tested load scenario classification model, and when abnormal electricity meter data occurs, timely alarms are issued or equipment parameters are adjusted in a timely manner.

Claims

1. A low power factor load scene classification method based on KAN network, characterized in that: include: Collect electricity meter data and time-align the data of each electricity meter to obtain a sample data set; The data of each electric energy meter in the sample data set is processed in sequence by removing abnormal data, filling in missing data and normalizing data to obtain a preprocessed data set; Perform feature extraction on each data in the preprocessed data set to obtain several groups of feature value samples, and integrate the feature value samples to obtain the electric energy meter data feature set; Mining the hidden correlation coefficients between the features in the electric energy meter data feature set; Perform feature crossover on two feature quantities whose hidden correlation coefficient is greater than the hidden correlation threshold to obtain new feature data; Using the new feature data, the CNN_KAN network load scenario classification model is trained to obtain a CNN_KAN network load scenario classification model that meets the performance requirements; The CNN_KAN network load scenario classification model that meets the performance requirements is deployed in the electricity meter software, and the load scenario classification is performed on the electricity meter data of the test scenario.

2. The low power factor load scenario classification method based on the KAN network according to claim 1 is characterized in that: The electric energy meter data includes voltage data, current data, active power data, power factor data and scene position data.

3. The low power factor load scenario classification method based on the KAN network according to claim 1 is characterized in that: The abnormal data elimination process adopts the box method; the missing data filling process adopts the polynomial function method.

4. The low power factor load scenario classification method based on the KAN network according to claim 1 is characterized in that: The characteristic value samples include time domain features and frequency domain features; the time domain features include mean value, standard deviation, skewness, kurtosis, maximum value, minimum value, peak-to-peak value, root mean square, rate of change, amplitude factor, waveform factor and impulse factor; the frequency domain features include rate spectrum density and spectrum centroid.

5. The low power factor load scenario classification method based on the KAN network according to claim 1 is characterized in that: The calculation of the hidden correlation coefficient between the features is specifically as follows: Based on the feature set of the electric energy meter data, the correlation coefficient between each feature and the load is calculated: in, For load and features The correlation coefficient between Characterized by No. eigenvalues; Characterized by The average eigenvalue of For the The load value corresponding to the eigenvalue sample; is the average load value; is the total number of eigenvalue samples; is the eigenvalue sample index; According to the correlation coefficient between each feature and the load, the hidden correlation coefficient between the features is calculated: in, Characterized by and features The hidden correlation coefficient between For load and features The correlation coefficient between Characterized by and features The correlation coefficient between For load and features The correlation coefficient between .

6. The low power factor load scenario classification method based on the KAN network according to claim 1 is characterized in that: The new feature data includes several new feature quantities: in, For the The new feature quantity is calculated by the eigenvalue samples; Characterized by No. dimensional latent vector; is the feature index; For the The first eigenvalue sample that participates in feature crossover eigenvalues.

7. The low power factor load scenario classification method based on the KAN network according to claim 1 is characterized in that: The CNN_KAN network load scenario classification model is specifically a CNN convolutional neural network; the fully connected layer of the CNN convolutional neural network is replaced by a KAN network.

Images