Power load identification method and system based on GAF and ResNet-CBAM fusion architecture

Abstract

This application discloses a method and system for identifying electric loads based on a GAF and ResNet‑CBAM fusion architecture, relating to the technical field of electric load identification. The method comprises collecting electric power data and storing it in a first database after initial data processing; performing event detection to determine the time of event occurrence; obtaining steady-state data before and after the event to generate electric quantity parameter data corresponding to the event and generate a signal waveform, and extracting characteristic harmonics; generating a two-dimensional feature image using a GAF algorithm, training the two-dimensional feature image using a ResNet‑CBAM neural network to generate a load type recognition model, and using the recognition model to identify the load type. The above scheme converts the original signal into an angular domain projection through improved GAF ​​encoding to achieve spatial representation of features. The improved GAF ​​can effectively convert electric load data into image form, making it easier for traditional neural networks to directly process it, thereby improving the accuracy of electric load identification.

Description

Technical Field

[0001] The present application relates to the technical field of power load identification, and in particular to a power load identification method and system based on a GAF and ResNet-CBAM fusion architecture. Background Art

[0002] With the rapid development of industrial automation, smart grids and new energy grid-connected technologies, problems such as frequent load fluctuations, increased nonlinear loads and malicious loads leading to overheating of wires, internal short circuits in electrical appliances and fires are becoming increasingly prominent.

[0003] Some existing power load identification systems struggle to extract effective distinguishing features from power data. This is due to two factors: precision errors during data collection, which lead to errors in subsequent load identification. Furthermore, the electrical characteristics of some loads are not distinct, and the electrical characteristics of different loads may be similar. These factors directly increase the difficulty of load identification in power systems. Therefore, a high-precision, intelligent load identification method is urgently needed to address these issues. Summary of the Invention

[0004] In response to the problem of low accuracy in power load identification in practical applications, the first purpose of this application is to provide a power load identification method based on the GAF and ResNet-CBAM fusion architecture, which converts the one-dimensional waveform data representing a specific load into an image through GAF technology, and then realizes the recognition and classification of specific image features through the fusion of the ResNet network and the CBAM module, and finally achieves a higher-precision and more intelligent load type identification effect. In order to implement the above-mentioned power load identification method, the second purpose of this application is to propose a power load identification system based on the GAF and ResNet-CBAM fusion architecture. At the same time, in order to facilitate the promotion and application of the above-mentioned power load identification method, the third purpose of this application is to propose a computer-readable storage medium, which is loaded with a program module for implementing the above-mentioned power load identification method. The specific plan is as follows:

[0005] A power load identification method based on the GAF and ResNet-CBAM fusion architecture includes:

[0006] After connecting a set load to the power line, power data of the power system is collected and acquired, and the data is stored in a first database after initial data processing;

[0007] Based on the power data in the first database, using a detection algorithm to perform event detection, determine the time of occurrence of each event and store it;

[0008] Obtain and calculate the power parameter data corresponding to each event based on the steady-state data before and after each event, and store the associated data in a second database;

[0009] Extracting electrical quantity parameter data associated with the event from the second database and generating a signal waveform, and extracting characteristic harmonics used to characterize the event using Fourier transform;

[0010] Performing image processing on the characteristic harmonics by using a GAF algorithm to generate a two-dimensional characteristic image;

[0011] Combined with the power parameter data stored in the second database, the generated two-dimensional feature image is trained using the ResNet-CBAM neural network to generate a load type recognition model;

[0012] Acquiring power data and identifying an output load type based on the load type identification model;

[0013] The events include the load entering a working state and the load ending a working state;

[0014] The electrical quantity parameter data includes current data and voltage data.

[0015] Through the above technical solution, the processed one-dimensional power data is converted into two-dimensional feature image data using the GAF algorithm, and then a recognition model for identifying the load type is obtained by training using a neural network. By combining the above load type recognition model with the collected power data, the load type working in the system can be accurately identified, thereby improving the accuracy and robustness of load identification.

[0016] Furthermore, collecting and acquiring power data of the power system and storing it in a first database after initial data processing includes:

[0017] The power data of the power system is collected through the mutual inductor and then output to the filter circuit to eliminate high-frequency noise;

[0018] Adjusting the amplitude of the current signal and / or voltage signal in the power data to adapt it to the input range of the analog-to-digital converter;

[0019] The power data is input into an analog-to-digital converter for analog-to-digital conversion to generate digitized power data and store it in a first database.

[0020] Through the above technical solution, power data can be converted into digital signals, eliminating noise interference and facilitating subsequent data processing steps.

[0021] Furthermore, event detection is performed using detection algorithms, including:

[0022] Aggregate power time series data is calculated based on the collected power data, and power difference series data is calculated after filtering and stored as input data for the RRCF algorithm.

[0023] Calculate the anomaly score of each data point in each power difference series data, and compare the calculated anomaly score with the set threshold to achieve event detection;

[0024] When the anomaly score is greater than a set threshold, it is determined that an event has occurred;

[0025] When the abnormality score is not greater than the set threshold, it is determined that no event has occurred.

[0026] Through the above technical solution, the power difference is used as input data to monitor abnormal data points in real time, so as to know when the event occurs, that is, when the load (electrical appliance) is turned on to start working or when it is turned off to end working; and the RRCF algorithm also has a self-checking processing effect. When a possible event is detected, the power difference threshold is used to further suppress false detection events.

[0027] Furthermore, the power parameter data corresponding to each event is calculated based on the steady-state data before and after each event. The calculation formula is:

[0028] V = (V off +V on ) / 2;

[0029] I=I on -I off ;

[0030] Wherein, V is the steady-state voltage of the power line, and V off is the voltage of the power line after the load finishes working, V on is the voltage of the power line after the load enters the working state, I is the steady-state current of the power line, and I on I is the current of the power line after the load enters the working state. off It is the current of the power line after the load finishes working;

[0031] The power parameter data of the above events obtained based on the steady-state data before and after the event also includes:

[0032] Based on the acquired power parameter data, the DDIM algorithm model is used to expand the data;

[0033] The expanded data is compared and evaluated with the original data to form the second database.

[0034] Through the above technical solution, not only can the power parameter data corresponding to the event, including voltage and current data, be accurately obtained, but the data can also be expanded, which is conducive to improving the training accuracy of subsequent recognition models.

[0035] Furthermore, the two-dimensional feature image is an RGB image, and generating the two-dimensional feature image by using the GAF algorithm includes:

[0036] Sampling characteristic harmonics to obtain harmonic sampling data as one-dimensional time series data;

[0037] The harmonic sampling data are normalized and represented using polar coordinates after PA transformation;

[0038] A single-channel image of time and frequency spectrum is obtained by encoding based on the GAF algorithm, and the single-channel image is corresponded to the three channels of the RGB color space to generate an RGB image.

[0039] Furthermore, the power load identification method further includes:

[0040] After the load is connected, power data of the power system is collected at different points of the power line, and stored as multiple power data groups in association with the sampling points;

[0041] Identifying data in a plurality of power data groups using the load type identification model and outputting a plurality of load type identification results;

[0042] Determine the recognition accuracy of each sampling point based on multiple load type recognition results;

[0043] Mark the identification accuracy of each sampling site.

[0044] Through the above technical solution, the best data sampling points in the power line can be found, which makes it easier for engineers to configure relevant power data sampling devices at the above sampling points and improve the accuracy of subsequent load type identification.

[0045] A power load identification system based on the GAF and ResNet-CBAM fusion architecture includes:

[0046] A data storage module configured to store setting data and provide a data interface for data retrieval or query;

[0047] The data acquisition module includes a data acquisition unit and a data initial processing unit, and is configured to acquire power data of the power system after the load is connected to the power line, and output the data to the data storage module after initial data processing to store it in the first database;

[0048] An event detection module is configured to be data-connected to the data storage module, and is used to detect events according to a set detection algorithm, determine the time of occurrence of each event, and output and store the time;

[0049] a data expansion module configured to be data-connected to the event detection module and the data storage module; after event detection is completed, the module determines the power parameter data of the event using the steady-state data before and after the event, and performs data expansion using the DDIM algorithm model; the expanded data is compared and evaluated with the original data, and then output to the data storage module to form a second database;

[0050] a characteristic waveform extraction module configured to extract electrical quantity parameter data associated with the event from the second database and generate a signal waveform, and to extract characteristic harmonics used to characterize the event using Fourier transform;

[0051] a feature image generation module configured to perform image processing on the feature harmonics using a GAF algorithm to generate a two-dimensional feature image;

[0052] a recognition model training module configured to use a ResNet-CBAM neural network to train the generated two-dimensional feature image in combination with the power parameter data stored in the second database to generate a load type recognition model;

[0053] a load type identification model configured to be data-connected to the data acquisition module, for acquiring power data and identifying an output load type based on the load type identification model;

[0054] The events include the load entering a working state and the load ending a working state;

[0055] The electrical quantity parameter data includes current data and voltage data.

[0056] Furthermore, the data acquisition unit includes a mutual inductor, which collects current and voltage signals in the power line at a set sampling point on the power line;

[0057] The data initial processing unit includes a signal filtering circuit unit, a signal amplitude modulation circuit unit, and an ADC converter, which is used to receive the current and voltage signals and perform filtering and noise reduction processing. The signal amplitude is then adjusted to a value suitable for the input range of the analog-to-digital converter by the signal amplitude modulation circuit unit and output to the ADC converter for conversion into a digital signal and then output.

[0058] Furthermore, the event detection module, data expansion module, characteristic waveform extraction module, and characteristic image generation module are all configured at the power system field end and are data-connected to the data acquisition module;

[0059] The data storage module includes a local memory and a cloud memory, and the recognition model training module is configured in the cloud server to receive data from the cloud data store for model training;

[0060] The load type identification model is deployed on the field side or the cloud server side, and is connected to the data acquisition module and the display device data configured on the field.

[0061] The above technical solution can facilitate the storage of power data, accelerate the training and optimization of the model, and facilitate the large-scale deployment and application of load type identification models.

[0062] A computer-readable storage medium is loaded with a program module. When the program module is executed by a processor, it is used to implement the power load identification method based on the GAF and ResNet-CBAM fusion architecture as described above.

[0063] The above technical solution is conducive to the promotion and use of the above power load identification method.

[0064] This application has at least one of the following beneficial effects:

[0065] (1) Through the improved GAF ​​encoding, the original signal can be converted into a projection in the angular domain, thereby realizing the spatial representation of the feature. In the load identification method of the present invention, the improved GAF ​​can effectively convert the power load data into an image form, so that the traditional CNN can directly process it.

[0066] (2) By introducing the ResNet-CBAM neural network, the key features of the load signal can be captured more accurately, thereby improving the recognition accuracy.

[0067] (3) Combining GAF with ResNet-CBAM can fully leverage the advantages of each. GAF converts payload signals into images, adapting to the input requirements of CNN; ResNet extracts complex features through its deep network structure; and CBAM further enhances the model's attention mechanism and improves its sensitivity to key features. This fusion architecture can better process the temporal and spatial features in payload signals, improving the accuracy and robustness of payload recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A schematic flow chart of the power load identification method of this application;

[0069] Figure 2 This is a schematic diagram of the functional module structure of the power load identification system of this application.

[0070] Figure numerals: 100, data storage module; 200, data acquisition module; 210, data acquisition unit; 220, data initial processing unit; 221, signal filtering circuit unit; 222, signal amplitude modulation circuit unit; 223, ADC converter; 300, event detection module; 400, data expansion module; 500, feature waveform extraction module; 600, feature image generation module; 700, recognition model training module; 800, load type recognition model; 900, data communication unit. DETAILED DESCRIPTION

[0071] The following describes the embodiments of the present application in detail, and examples of the embodiments are given in Figure 1-Figure 2 Shown in.

[0072] Throughout this specification, reference to the terms "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with the embodiment or example is included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0073] A power load identification method based on GAF and ResNet-CBAM fusion architecture, such as Figure 1 As shown, it mainly includes the following steps:

[0074] S100: After a set load is connected to a power line, power data of the power system is collected and acquired, and the data is stored as a first database after initial data processing.

[0075] In the embodiment of the present application, electric power data of the electric power system, including voltage and current data, is collected through mutual inductors installed on the power lines. The initial data step in the above-mentioned step S100 includes filtering and denoising the collected electric power data to eliminate high-frequency noise, and then adjusting the amplitude of the current signal and / or voltage signal in the electric power data to adapt it to the input range of the analog-to-digital converter, so as to facilitate input of the electric power data into the analog-to-digital converter for analog-to-digital conversion, generate digitized electric power data, and store it in the first database.

[0076] S200 , based on the power data in the first database, use a detection algorithm to perform event detection, determine the time when each event occurs, and store it.

[0077] In the implementation manner of the present application, the above-mentioned events include the load entering the working state and the load ending the working state. The present application scheme is used to detect the types of loads connected to the power system, including nonlinear loads, such as arc heating equipment, DC power equipment with AC rectification, switching power supply equipment, etc. The above-mentioned nonlinear power equipment, when connected to the power system as a harmonic source and working, will have an adverse effect on the quality of the grid voltage waveform.

[0078] In step S200, the Robust Random Cut Forest (RRCF) algorithm is mainly used to detect events. RRCF is an open source algorithm for anomaly detection in streaming data. The detection process includes:

[0079] S210, calculating aggregated power time series data based on the collected power data, and calculating power difference series data after filtering to serve as input data for the RRCF algorithm;

[0080] Aggregated power time series data analysis involves aggregating and analyzing chronologically ordered power data to understand its patterns and characteristics. Time series data includes power measurements recorded in chronological order, such as power consumption data per second or per minute, facilitating trend analysis and forecasting.

[0081] In step S210, first, mean pooling is performed on the aggregate power time series data with a high sampling rate to achieve the filtering effect. For the aggregate power time series data with a low sampling rate, since there is no drastic fluctuation, filtering is not required.

[0082] S220, calculating the anomaly score of each data point in each power difference sequence data, and comparing the calculated anomaly score with a set threshold to implement event detection:

[0083] When the anomaly score is greater than a set threshold, it is determined that an event has occurred;

[0084] When the abnormality score is not greater than the set threshold, it is determined that no event has occurred.

[0085] The RRCF algorithm uses power difference as input data and monitors abnormal data points in real time, thereby determining when an event occurs (i.e., when the appliance is turned on or off). The algorithm also features self-checking. When a possible event is detected, it uses a power difference threshold to further mitigate false positives.

[0086] S300 , obtaining and calculating the power parameter data corresponding to each event based on the steady-state data before and after each event, and storing the data in a related manner as a second database.

[0087] The electrical quantity parameter data includes current data and voltage data.

[0088] The power parameter data corresponding to each event is calculated based on the steady-state data before and after each event. The calculation formula is:

[0089] V = (V off +V on ) / 2; I=I on -I off Wherein, V is the steady-state voltage of the power line, and V off is the voltage of the power line after the load finishes working, V on is the voltage of the power line after the load enters the working state, I is the steady-state current of the power line, and I on I is the current of the power line after the load enters the working state. off It is the current in the power line after the load finishes working.

[0090] Acquiring the power parameter data of the above event based on the steady-state data before and after the event also includes: based on the acquired power parameter data, using the DDIM algorithm model to expand the data, and then comparing and evaluating the expanded data with the original data to form the second database.

[0091] The above-mentioned DDIM algorithm model refers to the Denoising Diffusion Implicit Models (Denoising Diffusion Implicit Models), which is an improved diffusion model designed to accelerate the process of generating data.

[0092] S400 , extracting electrical quantity parameter data associated with an event from a second database and generating a signal waveform, and utilizing Fourier transform (FFT) to extract characteristic harmonics for characterizing the event.

[0093] S500 , performing image processing on the characteristic harmonics by using a GAF algorithm to generate a two-dimensional characteristic image.

[0094] The GAF algorithm is the Gramian Angular Field (GAF) method, an effective method for converting one-dimensional time series data into a two-dimensional image representation. It works by treating each data point in the one-dimensional time series data as a point in vector space, calculating the cosine of the angles between these points, and then mapping these cosine values ​​to the pixels of the two-dimensional image, thereby generating an image that can reflect the dynamic and periodic characteristics of the time series.

[0095] In the embodiment of the present application, the two-dimensional feature image is an RGB image, and generating the two-dimensional feature image by the GAF algorithm includes:

[0096] S510, sampling characteristic harmonics to obtain harmonic sampling data as one-dimensional time series data;

[0097] S520 , normalizing the harmonic sampling data, such as normalizing to [0, 1] to eliminate the influence of different dimensions on the results, and then performing polar coordinate transformation (PA transformation) to express the data using polar coordinates;

[0098] S530 , encoding based on the GAF algorithm to obtain a single-channel image of time and frequency spectrum, and mapping the single-channel image to three channels of the RGB color space to generate an RGB image.

[0099] In step S530, the step of generating an image based on the GAF algorithm includes constructing a Gram matrix based on the processed harmonic sampling data, treating the time series data as vectors in a vector space, and calculating the inner products between these vectors. The angles between the vectors at different times are calculated based on the Gram matrix and converted to values ​​between 0 and 1, thereby generating a new angle matrix. The angle matrix is ​​used as the pixel values ​​of the image to generate a two-dimensional image. Each pixel value in the image corresponds to the cosine value of the angle between different times in the time series data, thereby preserving the temporal information and dynamic characteristics of the data.

[0100] S600 , combining the power parameter data stored in the second database, and using the ResNet-CBAM neural network to train the generated two-dimensional feature image to generate a load type recognition model.

[0101] The process of training and generating a load type recognition model includes: first preprocessing the above-mentioned RGB images, then dividing the data in the second database into a training set and a validation set, using a cross-entropy loss function, an Adam or SGD optimizer, then inputting the image into the network and calculating the output, calculating the loss value, backpropagating to update the network parameters, and repeating multiple training cycles until the model converges.

[0102] S700: Acquire power data and identify the output load type based on the load type identification model.

[0103] In actual applications, due to factors such as environmental electromagnetic interference and load distribution location, the power data collected at each sampling point on the power line has different harmonic characteristics. Therefore, the power load identification method further includes:

[0104] S800, after connecting to the load, collect power data of the power system at different locations on the power line, and associate the data with the sampling locations and store them as multiple power data groups;

[0105] S810, identifying data in a plurality of power data groups using the load type identification model, and outputting a plurality of load type identification results;

[0106] S820, determining the recognition accuracy of each sampling point based on the multiple load type recognition results;

[0107] S830, marking the recognition accuracy of each sampling site.

[0108] By using steps S800-S830, the best data sampling location in the power line can be found, which facilitates engineers to configure relevant power data sampling devices at the sampling location, thereby improving the accuracy of subsequent load type identification.

[0109] In order to implement the above-mentioned power load identification method, the embodiment of the present application also discloses a power load identification system based on the GAF and ResNet-CBAM fusion architecture, combined with Figure 2 As shown, it mainly includes a data storage module 100, a data acquisition module 200, an event detection module 300, a data expansion module 400, a feature waveform extraction module 500, a feature image generation module 600, a recognition model training module 700 and a load type recognition model 800.

[0110] The data storage module 100 is configured to store setting data and provide a data interface for data retrieval or query. In the embodiment of the present application, in order to facilitate the storage and retrieval of power data, the data storage module 100 includes a local memory and a cloud memory, and data transmission interaction is realized through a data communication unit 900 provided between the field end and the cloud memory.

[0111] The data acquisition module 200 includes a data acquisition unit 210 and an initial data processing unit 220, configured to acquire power data from the power system after a load is connected to the power line. After initial data processing, the data is output to the data storage module 100 for storage in the first database. Specifically, the data acquisition unit 210 includes a transformer that collects current and voltage signals from the power line at a predetermined sampling point on the power line. The initial data processing unit 220 includes a signal filtering circuit unit 221, a signal amplitude modulation circuit unit 222, and an ADC converter 223. The signal is then filtered and noise-reduced by the signal amplitude modulation circuit unit 222, which then adjusts the signal amplitude to a value compatible with the input range of the analog-to-digital converter. The signal is then converted to a digital signal by the ADC converter 223 and output.

[0112] The event detection module 300 is configured to be data-connected to the data storage module 100 and is used to detect events according to a set detection algorithm, determine the time of each event occurrence, and output and store it. In the embodiment of the present application, the RRCF (Robust Random Forest Partition) algorithm is used for event detection.

[0113] The data expansion module 400 is configured to be data-connected with the event detection module 300 and the data storage module 100. When the event detection is completed, the steady-state data before and after the event are used to determine the electrical parameter data of the above event. The electrical parameter data includes current data and voltage data. The DDIM algorithm model is then used to expand the data. The expanded data is compared and evaluated with the original data and then output to the data storage module 100 to form a second database.

[0114] The characteristic waveform extraction module 500 is configured to extract the electrical quantity parameter data associated with the event from the second database and generate a signal waveform, and then use Fourier transform to extract characteristic harmonics used to characterize the above event.

[0115] The feature image generation module 600 has a built-in GAF algorithm unit, which performs image processing on the feature harmonics through the GAF algorithm to generate a two-dimensional feature image. In the embodiment of the present application, an RGB image is generated.

[0116] The recognition model training module 700 is configured to use the ResNet-CBAM neural network to train the generated two-dimensional feature image in combination with the power parameter data stored in the second database to generate a load type recognition model 800.

[0117] The load type identification model 800 is configured to be data-connected to the data acquisition module 200 for acquiring power data and identifying the output load type based on the load type identification model 800 .

[0118] Furthermore, the event detection module 300, data expansion module 400, characteristic waveform extraction module 500, and characteristic image generation module 600 are all deployed on-site in the power system, stored as program modules in specific equipment, such as a PC on-site, and connected to the data acquisition module 200. To increase model training speed and facilitate continuous optimization of the recognition model, the recognition model training module 700 is deployed on a cloud server and receives data from the cloud data store for model training.

[0119] The trained load type identification model 800 is deployed on-site or on a cloud server, and is connected to the data acquisition module 200 and the display device configured on-site to identify the load type in the power system based on the received power data. The above technical solution facilitates the large-scale deployment and application of the load type identification model 800 described in this application.

[0120] To facilitate the promotion and application of the solution of the present application, the embodiment of the present application also discloses a computer-readable storage medium on which a program module is loaded. When the above program module is executed by a processor, it is used to implement the power load identification method based on the GAF and ResNet-CBAM fusion architecture as described above.

[0121] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A power load identification method based on GAF and ResNet-CBAM fusion architecture, characterized in that: include: After connecting a set load to the power line, power data of the power system is collected and acquired, and the data is stored in a first database after initial data processing; Based on the power data in the first database, using a detection algorithm to perform event detection, determine the time of occurrence of each event and store it; Obtain and calculate the power parameter data corresponding to each event based on the steady-state data before and after each event, and store the associated data in a second database; Extracting electrical quantity parameter data associated with the event from the second database and generating a signal waveform, and extracting characteristic harmonics used to characterize the event using Fourier transform; Performing image processing on the characteristic harmonics by using a GAF algorithm to generate a two-dimensional characteristic image; Combined with the power parameter data stored in the second database, the generated two-dimensional feature image is trained using the ResNet-CBAM neural network to generate a load type recognition model; Acquiring power data and identifying an output load type based on the load type identification model; The events include the load entering a working state and the load ending a working state; The electrical quantity parameter data includes current data and voltage data; The power parameter data corresponding to each event is calculated based on the steady-state data before and after each event. The calculation formula is: V = (Voff + Von) / 2; I=Ion-Ioff; Wherein, V is the steady-state voltage of the power line, Voff is the voltage of the power line after the load stops working, Von is the voltage of the power line after the load enters the working state, I is the steady-state current of the power line, Ion is the current of the power line after the load enters the working state, and Ioff is the current of the power line after the load stops working; The power parameter data of the above events obtained based on the steady-state data before and after the event also includes: Based on the acquired power parameter data, the DDIM algorithm model is used to expand the data; Comparing and evaluating the expanded data with the original data to form the second database; The power load identification method further includes: After the load is connected, power data of the power system is collected at different points of the power line, and stored as multiple power data groups in association with the sampling points; Identifying data in a plurality of power data groups using the load type identification model and outputting a plurality of load type identification results; Determine the recognition accuracy of each sampling point based on multiple load type recognition results; Mark the identification accuracy of each sampling site.

2. The power load identification method according to claim 1, characterized in that: The power data of the power system is collected and acquired, and stored in a first database after initial data processing, including: The power data of the power system is collected through the mutual inductor and then output to the filter circuit to eliminate high-frequency noise; Adjusting the amplitude of the current signal and / or voltage signal in the power data to adapt it to the input range of the analog-to-digital converter; The power data is input into an analog-to-digital converter for analog-to-digital conversion to generate digitized power data and store it in a first database.

3. The power load identification method according to claim 1, characterized in that: Utilize detection algorithms for event detection, including: Aggregate power time series data is calculated based on the collected power data, and power difference series data is calculated after filtering and stored as input data for the RRCF algorithm. Calculate the anomaly score of each data point in each power difference series data, and compare the calculated anomaly score with the set threshold to achieve event detection; When the anomaly score is greater than a set threshold, it is determined that an event has occurred; When the abnormality score is not greater than the set threshold, it is determined that no event has occurred.

4. The power load identification method according to claim 1, characterized in that: The two-dimensional feature image is an RGB image, and generating the two-dimensional feature image by using the GAF algorithm includes: Sampling characteristic harmonics to obtain harmonic sampling data as one-dimensional time series data; The harmonic sampling data are normalized and represented using polar coordinates after PA transformation; A single-channel image of time and frequency spectrum is obtained by encoding based on the GAF algorithm, and the single-channel image is corresponded to the three channels of the RGB color space to generate an RGB image.

5. A power load identification system for implementing the power load identification method based on the GAF and ResNet-CBAM fusion architecture as described in any one of claims 1 to 4, characterized in that: include: A data storage module (100) configured to store setting data and provide a data interface for data retrieval or query; The data acquisition module (200) comprises a data acquisition unit (210) and a data initial processing unit (220), and is configured to acquire power data of the power system after the load is connected to the power line, and output the data to the data storage module (100) after initial data processing, and store the data in a first database; An event detection module (300) is configured to be data-connected to the data storage module (100) and is used to perform event detection according to a set detection algorithm, determine the time of occurrence of each event, and output and store the time; A data expansion module (400) is configured to be data-connected to the event detection module (300) and the data storage module (100); after event detection is completed, the steady-state data before and after the event are used to determine the power parameter data of the event, and the DDIM algorithm model is used to expand the data; the expanded data is compared and evaluated with the original data, and then output to the data storage module (100) to form a second database; A characteristic waveform extraction module (500) is configured to extract electrical quantity parameter data associated with an event from a second database and generate a signal waveform, and to extract characteristic harmonics used to characterize the event using Fourier transform; A characteristic image generation module (600) is configured to perform image processing on the characteristic harmonics using a GAF algorithm to generate a two-dimensional characteristic image; A recognition model training module (700) is configured to train the generated two-dimensional feature image using a ResNet-CBAM neural network in combination with the electrical quantity parameter data stored in the second database to generate a load type recognition model (800); A load type identification model (800) configured to be data-connected to the data acquisition module (200) for acquiring power data and identifying an output load type based on the load type identification model (800); The events include the load entering a working state and the load ending a working state; The electrical quantity parameter data includes current data and voltage data.

6. The power load identification system according to claim 5, characterized in that: The data acquisition unit (210) comprises a mutual inductor, which collects current and voltage signals in the power line at a set sampling point on the power line; The data initial processing unit (220) comprises a signal filtering circuit unit (221), a signal amplitude modulation circuit unit (222), and an ADC converter (223), and is used to receive the current and voltage signals and perform filtering and noise reduction processing, and then adjust the signal amplitude to an input range suitable for the analog-to-digital converter through the signal amplitude modulation circuit unit (222) and output the signal to the ADC converter (223) for conversion into a digital signal and output.

7. The power load identification system according to claim 5, characterized in that: The event detection module (300), the data expansion module (400), the characteristic waveform extraction module (500), and the characteristic image generation module (600) are all configured at the power system field end and are data-connected to the data acquisition module (200); The data storage module (100) includes a local memory and a cloud memory, and the recognition model training module (700) is configured in a cloud server and receives data from a cloud data server for model training; The load type identification model (800) is deployed on the field side or the cloud server side, and is connected to the data acquisition module (200) and the display device data configured on the field.

8. A computer-readable storage medium, characterized in that A program module is loaded thereon, and when the program module is executed by a processor, it is used to implement the power load identification method based on the GAF and ResNet-CBAM fusion architecture as described in any one of claims 1 to 4.

Images