A load identification method based on power signal similarity clustering

Abstract

The present invention relates to the technical field of power system electricity consumption monitoring and analysis, and discloses a load identification method based on power signal similarity clustering. A collector or a special transformer terminal deployed at the user side incoming line is used to perform denoising on the collected voltage and current waveform data using a wavelet filtering algorithm. Time domain features, frequency domain features and time-frequency domain features are extracted from the pre-processed voltage and current waveform data respectively, and a measurement method combining Euclidean distance and cosine similarity is used. The present invention uses the characteristic similarity of voltage and current waveforms to perform non-invasive electrical equipment identification, which is deployed on a collector or a special transformer terminal to realize energy consumption structure analysis on the user side. With the energy consumption structure analysis, users can clearly know the energy consumption status and usage patterns of each electrical equipment, thereby formulating a more scientific and reasonable electricity consumption plan, avoiding the use of high-energy-consuming equipment during peak hours, reducing unnecessary electricity waste, and thus significantly reducing electricity costs.

Description

Technical Field

[0001] The present invention relates to the technical field of power system electricity consumption monitoring and analysis, and in particular to a load identification method based on power signal similarity clustering. Background Art

[0002] Power signal similarity clustering is a load identification method. It first collects voltage and current waveform data at high frequency and for a long time. After preprocessing, it extracts features from multiple domains. It measures feature similarity by combining Euclidean distance and cosine similarity, and then uses the DBSCAN algorithm to adaptively adjust parameters for clustering.

[0003] Traditional methods for identifying electrical devices typically require installing individual monitoring sensors on each device. This is not only costly and complex to install, but also hinders large-scale application. In recent years, non-invasive load identification technology has become a research hotspot. It analyzes the voltage and current signals at the user's main incoming line to identify electrical devices. However, existing non-invasive load identification methods lack accuracy and reliability when dealing with complex power consumption environments and multiple devices operating simultaneously. In particular, many issues remain in signal feature extraction and cluster analysis, making it difficult to accurately distinguish different types of electrical devices. Summary of the Invention

[0004] The purpose of the present invention is to provide a load identification method based on power signal similarity clustering to overcome the shortcomings of the existing technology. The method uses the characteristic similarity of voltage or current waveforms to perform non-invasive electrical equipment identification and is deployed on a collector or dedicated transformer terminal to achieve accurate analysis of the energy consumption structure on the user side.

[0005] The present invention provides the following technical solutions:

[0006] A load identification method based on power signal similarity clustering includes the following steps:

[0007] S1. Data collection steps: Use a collector or dedicated transformer terminal deployed at the user-side incoming line to synchronously collect voltage and current waveform data at a frequency of no less than 1000 sampling points per second, and continuously collect data for no less than 24 hours. Accurately record the timestamp of each sampling point.

[0008] S2. Waveform preprocessing step: Use the wavelet filtering algorithm to denoise the collected voltage and current waveform data. Then use the maximum-minimum normalization method to normalize the amplitude of the denoised waveform data to eliminate the influence of power differences between different devices.

[0009] S3, feature extraction step: extracting time domain features, frequency domain features, and time-frequency domain features from the preprocessed voltage and current waveform data, respectively, wherein the time domain features include at least mean, effective value, variance, and crest factor, the frequency domain features are obtained by fast Fourier transform, and the time-frequency domain features are extracted by wavelet packet transform;

[0010] S4. Similarity measurement steps: using Euclidean distance and cosine similarity The combined measurement method is used to calculate the similarity between the feature vectors corresponding to different devices, and the final similarity measurement value S is obtained by weighted average. The specific formula is S=ɑ×(1- )+(1-ɑ)×cos(θ), where d(x,y) is the Euclidean distance, is the maximum value of the Euclidean distance between all pairs of feature vectors, cos(θ) is the cosine similarity, and α is the weight coefficient ranging from 0 to 1;

[0011] S5. Cluster analysis step: Use DBSCAN density-based spatial clustering algorithm to perform cluster analysis on the similarity measurement results, and preset the neighborhood radius and minimum points Parameters, and dynamically adjust the preset neighborhood radius according to the distribution of data during the clustering process and minimum points ;

[0012] S6. Device identification and energy usage structure analysis steps: Build a database containing feature templates of common electrical equipment, compare the clusters obtained by clustering with the features in the template database, and determine the type of electrical equipment corresponding to each cluster based on the degree of match; count the proportion of different equipment types in the total power usage time and power consumption, and then analyze the energy usage structure on the user side to provide users with energy-saving suggestions and power optimization solutions.

[0013] Preferably, the denoising process of the collected voltage and current waveform data using a wavelet filtering algorithm is specifically as follows: first, a wavelet decomposition operation is performed on the waveform data to decompose the waveform data into different frequency bands and scales to obtain sub-signals of different scales and frequencies; threshold processing is performed on each sub-signal to select a suitable threshold to remove the part containing noise, and the specific threshold can be determined based on experience or statistical methods; finally, wavelet reconstruction is performed to recombine the processed sub-signals to obtain the denoised waveform data.

[0014] Preferably, by formula = Calculate the mean using the formula RMS= Calculate the effective value by the formula Var= Calculate the variance by the formula, CF= Calculate the crest factor, where is the value of the signal at the i-th sampling point, and N is the number of sampling points.

[0015] Preferably, when extracting the frequency domain features, components in the frequency range of 0-1000 Hz are taken to analyze the amplitude and phase information of each frequency.

[0016] Preferably, the weight coefficient α is determined through multiple experimental optimizations. By setting different α values, the accuracy, stability and matching degree of the clustering results with the device feature template are evaluated, and finally the α value that performs best under the experimental evaluation indicators is selected as the weight coefficient. The specific value will vary depending on different scenarios and experimental conditions.

[0017] Preferably, the initial value of the preset neighborhood radius is set to 0.2, and the minimum number of points is The initial value is set to 5, and it is dynamically adjusted during the clustering process. and The method is: count the density distribution of the current cluster, if the density of the current cluster is higher than the preset density threshold, then reduce it appropriately To refine the clustering results; if the density of the current cluster is lower than the preset density threshold, increase it appropriately To merge more sample points, and adjust according to the changes in the number of samples in the cluster .

[0018] Preferably, when the device is identified, when the matching degree between the cluster and the template database features is less than 80%, the cluster is marked as an unknown device cluster; for the unknown device cluster, the features in its feature vector are first manually analyzed to infer the device type from the waveform shape and power changes; if manual analysis is difficult to determine, support vector machine SVM and random forest machine learning models are further used for training and identification, and the model is trained using feature data of known types of devices, and then the feature data of the unknown device cluster is input into the model for prediction.

[0019] Preferably, in the energy consumption structure analysis, detailed statistics are made on the energy consumption ratio and usage time distribution of different devices. In addition to counting the proportion of different equipment types in the total power consumption time and power consumption, statistics are also broken down by time period, and the energy consumption and usage of equipment in different time periods on weekdays and weekends, daytime and nighttime are counted; a detailed energy consumption structure report is generated based on these statistical data, which includes analysis of peak and low period of equipment use, analysis of equipment with higher energy consumption, and energy-saving suggestions for different equipment.

[0020] Preferably, in the data collection step, while collecting voltage and current waveform data, ambient temperature and humidity environmental parameters are also collected, because environmental factors may affect the operating status and power signal characteristics of electrical equipment. When performing feature extraction and cluster analysis later, these environmental parameters are considered as auxiliary features to improve the accuracy of load identification.

[0021] Preferably, the database of common electrical equipment feature templates will be updated and maintained regularly, and the updating method is: collecting feature data of newly emerging electrical equipment and adding it to the database; for existing equipment feature templates, regularly re-collecting data to update features based on actual usage and changes in equipment performance; at the same time, using big data analysis technology to optimize the feature data in the database, remove possible error data and redundant data, and improve the quality and availability of the database.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] The present invention uses the characteristic similarity of voltage and current waveforms to perform non-invasive identification of electrical equipment, and is deployed on a collector or a dedicated transformer terminal to realize energy consumption structure analysis on the user side. With the energy consumption structure analysis, users can clearly know the energy consumption status and usage patterns of each electrical equipment, so as to formulate a more scientific and reasonable electricity consumption plan, avoid the use of high-energy-consuming equipment during peak hours, reduce unnecessary electricity waste, and thus significantly reduce electricity costs, achieving a dual improvement in energy conservation and emission reduction and economic benefits. Continuous monitoring of voltage or current waveform characteristics can detect abnormal operating conditions of electrical equipment. When a device fails or its performance degrades, its power waveform will change. By detecting these signals in advance, users can perform equipment maintenance and inspection in a timely manner, which can not only avoid further deterioration of equipment failures and extend the service life of the equipment, but also prevent safety accidents caused by equipment failures, and ensure the safety and stability of homes or workplaces. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Flowchart of the entire invention. DETAILED DESCRIPTION

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0026] See also Figure 1 , a load identification method based on power signal similarity clustering, comprising the following steps:

[0027] S1. Data collection steps: Use a collector or dedicated transformer terminal deployed at the user-side incoming line to synchronously collect voltage and current waveform data at a frequency of no less than 1000 sampling points per second, and continuously collect data for no less than 24 hours. Accurately record the timestamp of each sampling point.

[0028] S2. Waveform preprocessing step: Use the wavelet filtering algorithm to denoise the collected voltage and current waveform data. Then use the maximum-minimum normalization method to normalize the amplitude of the denoised waveform data to eliminate the influence of power differences between different devices.

[0029] S3, feature extraction step: extracting time domain features, frequency domain features, and time-frequency domain features from the preprocessed voltage and current waveform data, respectively, wherein the time domain features include at least mean, effective value, variance, and crest factor, the frequency domain features are obtained by fast Fourier transform, and the time-frequency domain features are extracted by wavelet packet transform;

[0030] S4. Similarity measurement step: Use the measurement method combining Euclidean distance and cosine similarity d to calculate the similarity between the feature vectors corresponding to different devices, and obtain the final similarity measurement value S by weighted average. The specific formula is S=ɑ×(1- )+(1-ɑ)×cos(θ), where d(x,y) is the Euclidean distance, which is the maximum value of the Euclidean distance between all pairs of feature vectors, cos(θ) is the cosine similarity, and α is the weight coefficient ranging from 0 to 1;

[0031] S5. Cluster analysis step: Use DBSCAN density-based spatial clustering algorithm to perform cluster analysis on the similarity measurement results, and preset the neighborhood radius and minimum points Parameters, and dynamically adjust the preset neighborhood radius according to the distribution of data during the clustering process and minimum points ;

[0032] S6. Device Identification and Energy Usage Structure Analysis: Build a database containing feature templates of common electrical devices. Compare the features of each cluster obtained from clustering with the features in the template database. Determine the type of electrical device corresponding to each cluster based on the degree of match. Count the proportion of different device types in total power usage time and power consumption, and then analyze the energy usage structure on the user side to provide users with energy-saving suggestions and power optimization solutions.

[0033] Furthermore, the denoising process of the collected voltage and current waveform data using the wavelet filtering algorithm is specifically as follows: first, a wavelet decomposition operation is performed on the waveform data to decompose the waveform data into different frequency bands and scales to obtain sub-signals of different scales and frequencies; threshold processing is performed on each sub-signal to select an appropriate threshold to remove the part containing noise, and the specific threshold can be determined based on experience or statistical methods; finally, wavelet reconstruction is performed to recombine the processed sub-signals to obtain denoised waveform data;

[0034] By formula = Calculate the mean using the formula RMS= Calculate the effective value by the formula Var= Calculate the variance by the formula, CF= Calculate the crest factor, where is the value of the signal at the i-th sampling point, and N is the number of sampling points;

[0035] Furthermore, when extracting frequency domain features, the components with a frequency range of 0-1000Hz are taken to analyze the amplitude and phase information of each frequency. The weight coefficient α is determined through multiple experimental optimizations. By setting different α values, the accuracy, stability and matching index of the clustering results with the device feature template are evaluated. Finally, the α value that performs best under the experimental evaluation index is selected as the weight coefficient, and the specific value will vary depending on different scenarios and experimental conditions. The initial value of the preset neighborhood radius is set to 0.2, and the minimum number of points is 0. The initial value is set to 5, and it is dynamically adjusted during the clustering process. and The method is: count the density distribution of the current cluster, if the density of the current cluster is higher than the preset density threshold, then reduce it appropriately To refine the clustering results; if the density of the current cluster is lower than the preset density threshold, increase it appropriately To merge more sample points, and adjust according to the changes in the number of samples in the cluster ;

[0036] Furthermore, when identifying devices, if the matching degree between a cluster and the template database features is less than 80%, the cluster is marked as an unknown device cluster. For unknown device clusters, the features in their feature vectors are manually analyzed to infer the device type based on waveform shape and power changes. If manual analysis is difficult to determine, support vector machines (SVMs) and random forest machine learning models are used for training and identification. The models are trained using feature data of known device types, and then the feature data of unknown device clusters is input into the models for prediction.

[0037] The energy usage structure analysis includes detailed statistics on the energy consumption ratio and usage time distribution of different devices. In addition to counting the proportion of different device types in total power usage time and power consumption, statistics are also broken down by time period, including the energy consumption and usage of devices during weekdays and weekends, and during the day and night. Based on these statistics, a detailed energy usage structure report is generated, which includes an analysis of peak and off-peak periods for device usage, an analysis of devices with high energy consumption, and energy-saving recommendations for different devices.

[0038] Furthermore, in the data collection step, while collecting voltage and current waveform data, environmental parameters such as ambient temperature and humidity are also collected, because environmental factors may affect the operating status and power signal characteristics of electrical equipment. These environmental parameters can be considered as auxiliary features during subsequent feature extraction and cluster analysis to improve the accuracy of load identification. The database of common electrical equipment feature templates will be regularly updated and maintained. The update method is: feature data of newly emerging electrical equipment is collected and added to the database; for existing equipment feature templates, data is regularly re-collected to update features based on actual usage and changes in equipment performance; at the same time, big data analysis technology is used to optimize the feature data in the database, remove possible error data and redundant data, and improve the quality and availability of the database;

[0039] The overall process of the present invention is:

[0040] Data collection phase: Install a collector or dedicated transformer terminal at the user's incoming line end, sampling at [X] points per second. Adjust according to actual conditions. It is recommended to select a sampling frequency that can accurately capture subtle changes in the signal. For example, if X = 1000, the sampling frequency of 1000 sampling points per second will synchronously collect voltage and current waveform data.

[0041] Continuously record for at least [Y] hours, where Y = 24 hours, to ensure that waveform data of electrical equipment at different usage periods is covered, and accurately record the timestamp of each sampling point;

[0042] Waveform preprocessing stage: The collected waveform data is denoised using the wavelet filtering algorithm. Wavelet filtering can perform targeted processing based on the different frequency components of the signal, effectively removing high-frequency noise while retaining the main features of the signal. The specific steps are: first, wavelet decomposition is performed on the waveform data to decompose it into sub-signals of different scales and frequencies; then, threshold processing is performed on each sub-signal to remove the part corresponding to the noise; finally, wavelet reconstruction is performed to obtain the denoised waveform data;

[0043] In order to eliminate the amplitude influence caused by the power difference of different devices, the maximum-minimum normalization method is used to normalize the denoised voltage and current waveforms. The specific formula is as follows:

[0044] ;

[0045] Where x is the initial data, and are the minimum and maximum values ​​in the data sequence, respectively. is the normalized data;

[0046] Calculate the mean, RMS value, variance, and crest factor of the waveform. The mean reflects the average level of the signal, the RMS value is related to the actual power of the device, the variance reflects the degree of signal fluctuation, and the crest factor reflects the peak characteristics of the waveform. The specific calculation formula is as follows:

[0047] Mean: = ;

[0048] Effective value: RMS= ;

[0049] Variance: Var=

[0050] Crest Factor: CF= ;

[0051] in is the value of the signal at sampling point i, and N is the number of sampling points;

[0052] Use Fast Fourier Transform (FFT) to convert the time domain waveform into a frequency domain signal, extract the amplitude and phase information of each frequency component in the frequency range of 0-1000Hz, and analyze it to capture the main harmonic characteristics of the device;

[0053] Wavelet packet transform is used to extract time-frequency domain features. Wavelet packet transform can perform more refined time-frequency decomposition of signals, provide richer feature information, and describe transient changes during equipment operation by calculating the energy and entropy characteristics of different wavelet packet nodes.

[0054] The similarity between feature vectors of different devices is calculated by combining Euclidean distance and cosine similarity. Euclidean distance can intuitively reflect the distance between feature vectors in space, while cosine similarity focuses more on the directional similarity of feature vectors.

[0055] The calculation formula is as follows:

[0056] Euclidean distance: d(s,y) = ;

[0057] Cosine similarity: = =

[0058] Where x=( ... ), and y=( ... ) are the feature vectors of the two devices;

[0059] Taking into account the calculation results of Euclidean distance and cosine similarity, the final similarity measurement value S is obtained by weighted average:

[0060] S= ;

[0061] in is the maximum value of the Euclidean distance between all pairs of eigenvectors, is the weight coefficient (0< <1, can be determined through experimental optimization, );

[0062] Select DBSCAN, a density-based spatial clustering algorithm for cluster analysis. The DBSCAN algorithm can handle clusters of any shape and is robust to noisy data. Set a suitable neighborhood radius. and minimum points Parameters, where =0.2, =5, the similarity measurement result is used as input for clustering operation;

[0063] In order to improve the accuracy and reliability of clustering, an adaptive adjustment strategy is introduced in the clustering process to dynamically adjust the and Parameters to ensure the stability of clustering results;

[0064] Establish a database containing feature templates of common electrical devices. Match each cluster obtained by clustering with the features in the template database. Determine the device type corresponding to the cluster based on the template with the highest matching degree. For clusters with a matching degree below a set threshold (e.g., 80%), mark them as unknown device clusters and use manual analysis and further machine learning model training to determine their type.

[0065] Statistics are collected on the proportion of each device type in the total power consumption time and power consumption. Based on the clustering and identification results, the energy consumption ratio and usage time distribution parameters of different devices are calculated. By analyzing the above result parameters, a detailed energy consumption structure report is provided to users, and targeted energy-saving suggestions and power optimization plans are proposed.

[0066] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and alterations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A load identification method based on power signal similarity clustering, characterized in that: The following steps are involved: S1. Data collection steps: Use a collector or dedicated transformer terminal deployed at the user-side incoming line to synchronously collect voltage and current waveform data at a frequency of no less than 1000 sampling points per second, and continuously collect data for no less than 24 hours. Accurately record the timestamp of each sampling point. S2. Waveform preprocessing step: Use the wavelet filtering algorithm to denoise the collected voltage and current waveform data. Then use the maximum-minimum normalization method to normalize the amplitude of the denoised waveform data to eliminate the influence of power differences between different devices. S3, feature extraction step: extracting time domain features, frequency domain features, and time-frequency domain features from the preprocessed voltage and current waveform data, respectively, wherein the time domain features include at least mean, effective value, variance, and crest factor, the frequency domain features are obtained by fast Fourier transform, and the time-frequency domain features are extracted by wavelet packet transform; S4. Similarity measurement steps: using Euclidean distance and cosine similarity The combined measurement method calculates the similarity between the feature vectors corresponding to different devices, and obtains the final similarity measurement value S by weighted average. The specific formula is + (1-ɑ) × cos (θ), where d (x, y) is the Euclidean distance, is the maximum value of the Euclidean distance between all pairs of feature vectors, cos(θ) is the cosine similarity, and α is the weight coefficient ranging from 0 to 1; S5. Cluster analysis step: Use DBSCAN density-based spatial clustering algorithm to perform cluster analysis on the similarity measurement results, and preset the neighborhood radius and minimum points Parameters, and dynamically adjust the preset neighborhood radius according to the distribution of data during the clustering process and minimum points ; S6. Device identification and energy usage structure analysis steps: Build a database containing feature templates of common electrical equipment, compare the clusters obtained by clustering with the features in the template database, and determine the type of electrical equipment corresponding to each cluster based on the degree of match; count the proportion of different equipment types in the total power usage time and power consumption, and then analyze the energy usage structure on the user side to provide users with energy-saving suggestions and power optimization solutions.

2. The load identification method based on power signal similarity clustering according to claim 1, characterized in that: The denoising process of the collected voltage and current waveform data using the wavelet filtering algorithm is specifically as follows: first, a wavelet decomposition operation is performed on the waveform data to decompose the waveform data into different frequency bands and scales to obtain sub-signals of different scales and frequencies; threshold processing is performed on each sub-signal to select an appropriate threshold to remove the part containing noise, and the specific threshold can be determined based on experience or statistical methods; finally, wavelet reconstruction is performed to recombine the processed sub-signals to obtain denoised waveform data.

3. The load identification method based on power signal similarity clustering according to claim 1, characterized in that: By formula = Calculate the mean using the formula RMS= Calculate the effective value by the formula Var= Calculate the variance by the formula, CF= Calculate the crest factor, where is the value of the signal at the i-th sampling point, and N is the number of sampling points.

4. The load identification method based on power signal similarity clustering according to claim 1, characterized in that: When extracting the frequency domain features, the components in the frequency range of 0-1000 Hz are taken to analyze the amplitude and phase information of each frequency.

5. The load identification method based on power signal similarity clustering according to claim 1, characterized in that: The weight coefficient α is determined through multiple experimental optimizations. By setting different α values, the accuracy, stability, and matching degree of the clustering results with the device feature template are evaluated. Finally, the α value that performs best under the experimental evaluation indicators is selected as the weight coefficient. The specific value will vary depending on different scenarios and experimental conditions.

6. The load identification method based on power signal similarity clustering according to claim 1, characterized in that: The preset neighborhood radius The initial value is set to 0.2, and the minimum number of points The initial value is set to 5, and it is dynamically adjusted during the clustering process. and The method is: count the density distribution of the current cluster, if the density of the current cluster is higher than the preset density threshold, then reduce it appropriately To refine the clustering results; if the density of the current cluster is lower than the preset density threshold, increase it appropriately To merge more sample points, and adjust according to the changes in the number of samples in the cluster .

7. The load identification method based on power signal similarity clustering according to claim 1, characterized in that: During the device identification process, when the matching degree between the cluster and the template database features is less than 80%, the cluster is marked as an unknown device cluster. For unknown device clusters, the features in their feature vectors are first manually analyzed to infer the device type from the waveform shape and power changes. If manual analysis is difficult to determine, support vector machines (SVMs) and random forest machine learning models are further used for training and identification. The model is trained using feature data of known types of devices, and then the feature data of the unknown device cluster is input into the model for prediction.

8. The load identification method based on power signal similarity clustering according to claim 1, characterized in that: In the energy consumption structure analysis, detailed statistics are made on the energy consumption ratio and usage time distribution of different devices. In addition to counting the proportion of different equipment types in the total power consumption time and power consumption, statistics are also broken down by time period, and the energy consumption and usage of equipment in different time periods such as weekdays and weekends, daytime and nighttime are counted; based on these statistical data, a detailed energy consumption structure report is generated, which includes an analysis of peak and low period of equipment use, an analysis of equipment with a higher energy consumption ratio, and energy-saving suggestions for different equipment.

9. The load identification method based on power signal similarity clustering according to claim 1, characterized in that: In the data collection step, while collecting voltage and current waveform data, ambient temperature and humidity parameters are also collected, because environmental factors may affect the operating status and power signal characteristics of electrical equipment. When performing feature extraction and cluster analysis later, these environmental parameters are considered as auxiliary features to improve the accuracy of load identification.

10. The load identification method based on power signal similarity clustering according to claim 1, characterized in that: The database of common electrical equipment feature templates will be updated and maintained regularly. The updating method is: collecting feature data of newly emerging electrical equipment and adding it to the database; for existing equipment feature templates, regularly re-collecting data to update features based on actual usage and changes in equipment performance; at the same time, using big data analysis technology to optimize the feature data in the database, remove possible error data and redundant data, and improve the quality and availability of the database.

Images