A load feature extraction and identification method, system, device and medium against voltage change

Abstract

The application discloses a load feature extraction and identification method, system, device and medium against voltage change, belonging to the technical field of power distribution load identification, comprising collecting steady-state voltage waveform and steady-state current waveform, preprocessing the waveforms, and constructing a load waveform database; calculating multi-category load waveform features from the voltage waveform and the current waveform to form a feature vector; grouping and counting the feature vector, implementing joint discrimination based on a multi-statistic set integration criterion to obtain a feature subset; performing data enhancement processing on the voltage waveform and the current waveform to generate an enhanced sample set and construct a neural network load identification model; calculating a real-time feature vector of the feature subset according to real-time voltage waveform and real-time current waveform, inputting the real-time feature vector into the neural network load identification model, and outputting a real-time load identification result. The application can provide stable and accurate load identification in a power system with large voltage fluctuation, and provide reliable support for load side management.

Description

Technical Field

[0001] This invention relates to the field of power distribution load identification technology, specifically to a method, system, device, and medium for extracting and identifying load characteristics resistant to voltage changes. Background Technology

[0002] With the continuous development and increasing intelligence of power systems, power load identification has become a crucial aspect of power load-side management. In recent years, with the rapid development of artificial intelligence and deep learning technologies, neural network-based load identification methods have received widespread attention. By learning more complex feature representations, the accuracy of load identification can be improved to some extent. However, existing neural network load identification models typically rely on static voltage waveform data for training, failing to consider the actual impact of voltage fluctuations on the identification results, leading to model instability under conditions of significant voltage variations. Furthermore, existing training methods often cannot effectively avoid the problem of excessive feature dispersion caused by voltage changes when dealing with load characteristics under multiple voltage conditions. Summary of the Invention

[0003] In view of the above-mentioned problems, the present invention provides a method, system, device and medium for extracting and identifying load features resistant to voltage changes.

[0004] Therefore, the technical problems solved by this invention are: how to address the fact that existing load identification methods do not consider the impact of voltage changes on load waveforms, resulting in the inability to accurately extract load features under voltage fluctuation conditions, leading to poor accuracy and stability of identification results; existing load identification methods lack an effective feature screening mechanism under multi-voltage conditions, resulting in the inability to screen out stable load features at different voltage points, thus affecting the performance of the model and the accuracy of the identification results; and existing methods fail to effectively handle the cross-voltage consistency problem between voltage and current waveforms when training the load identification model, resulting in unstable output distribution of the load identification model at different voltage points, affecting the model's performance in practical applications.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for extracting and identifying load features resistant to voltage variations, comprising: acquiring steady-state voltage waveforms and steady-state current waveforms at multiple target voltage points; preprocessing the waveforms to construct a load waveform database under different voltage operating conditions; calculating multi-category load waveform features from the voltage and current waveforms to form feature vectors; performing grouped statistical analysis on the feature vectors and implementing joint discrimination based on a multi-statistic integration criterion to obtain feature subsets; performing data augmentation processing on the voltage and current waveforms based on the load waveform database to generate an enhanced sample set and constructing a neural network load identification model; calculating real-time feature vectors of the feature subsets based on real-time voltage and current waveforms; inputting the real-time feature vectors into the neural network load identification model to output real-time load identification results.

[0006] As a preferred embodiment of the load feature extraction and identification method for resisting voltage changes described in this invention, the step of performing data augmentation processing on voltage and current waveforms based on a load waveform database to generate an augmented sample set and constructing a neural network load identification model includes: introducing a data augmentation training strategy; performing cross-voltage resampling consistency, noise injection, and phase perturbation data augmentation processing on voltage and current waveforms based on a load waveform database to generate an augmented sample set; using the feature vectors of the feature subset as input to the neural network load identification model and using the load category as a supervision label for training; and using a loss function with cross-voltage consistency constraints to constrain the output distribution of the same load at different voltage points during the training process to obtain a neural network load identification model that resists voltage changes.

[0007] This preferred solution uses data augmentation and cross-voltage consistency constraint training to enable the neural network load identification model to adapt to different voltage operating conditions.

[0008] As a preferred embodiment of the load feature extraction and identification method for resisting voltage changes described in this invention, the step of acquiring steady-state voltage waveforms and steady-state current waveforms at multiple target voltage points, preprocessing the waveforms, and constructing a load waveform database under different voltage operating conditions includes: supplying power to the load based on multiple target voltage points within the voltage adjustment range of the adjustable power supply; synchronously acquiring steady-state voltage waveforms and steady-state current waveforms at each target voltage point using voltage and current sensors; performing time alignment, noise reduction filtering, and power frequency cycle segmentation on the acquired waveforms; and associating and storing the processed waveforms with corresponding voltage identifiers and load category identifiers to form a load waveform database under different voltage operating conditions.

[0009] As a preferred embodiment of the load feature extraction and identification method for resisting voltage changes described in this invention, the step of calculating multiple categories of load waveform features from voltage waveforms and current waveforms to form a feature vector includes: for each record in the load waveform database, calculating multiple categories of load waveform features based on voltage waveforms and current waveforms within the same power frequency cycle; and forming a feature vector from each category of features in a preset order.

[0010] As a preferred embodiment of the load feature extraction and identification method for resisting voltage changes described in this invention, the step of grouping and statistically analyzing the feature vectors and performing joint discrimination based on a multi-statistic integration criterion to obtain a feature subset includes: grouping and statistically analyzing the feature vectors according to load categories; calculating the cross-voltage dispersion index and intra-class consistency index for each feature within the voltage regulation range; performing joint discrimination on the dispersion index and consistency index of each feature based on a preset multi-statistic integration criterion; eliminating features whose dispersion with voltage change exceeds a threshold; and obtaining a feature subset whose variance with voltage change satisfies the threshold constraint.

[0011] This preferred scheme uses a combination of cross-voltage dispersion index and intra-class consistency index to screen feature subsets, enabling them to have stability against voltage changes.

[0012] As a preferred embodiment of the load feature extraction and identification method for resisting voltage changes described in this invention, the step of calculating the real-time feature vector of the feature subset based on the real-time voltage waveform and the real-time current waveform, inputting the real-time feature vector into the neural network load identification model, and outputting the real-time load identification result includes: acquiring real-time voltage waveform and real-time current waveform by voltage sensor and current sensor; performing preprocessing on the real-time waveform; calculating the real-time feature vector of the feature subset; inputting the real-time feature vector into the neural network load identification model for resisting voltage changes, and outputting the real-time load identification result.

[0013] This preferred solution ensures that the online recognition input remains consistent with the training input by applying the same preprocessing to the real-time waveform as during the database construction phase.

[0014] As a preferred embodiment of the load feature extraction and identification method for resisting voltage changes described in this invention, the following steps are taken: for each record in the load waveform database, multiple categories of load waveform features are calculated based on the voltage and current waveforms within the same power frequency cycle, including: extracting voltage cycle segment vectors and current cycle segment vectors within the same power frequency cycle; calculating multiple categories of load waveform features, including time-domain statistical features, amplitude and phase features, harmonic spectrum features, power and energy features, and waveform shape features; calculating time-domain statistical features based on the voltage cycle segment vectors and current cycle segment vectors, including mean, variance, peak-to-peak value, and peak factor; and calculating time-domain statistical features based on the power supply base... The amplitude and phase characteristics are calculated using discrete Fourier projection with frequency synchronization. These characteristics include the fundamental voltage amplitude, fundamental current amplitude, fundamental voltage phase, fundamental current phase, and fundamental phase difference. Harmonic spectrum characteristics are calculated based on a preset harmonic order. These characteristics include the voltage harmonic amplitude, current harmonic amplitude, total harmonic distortion (THD) of the voltage, and total harmonic distortion (THD) of the current. Power and energy characteristics are calculated based on an instantaneous power sequence. These characteristics include active power, apparent power, reactive power, and power frequency cycle energy. Waveform shape characteristics are calculated based on voltage and current sampling points after removing the DC component. These characteristics include normalized cross-correlation peak value, waveform asymmetry, and kurtosis.

[0015] This invention provides a load feature extraction and identification system that is resistant to voltage changes.

[0016] To address the aforementioned technical problems, this invention provides the following technical solution: a load feature extraction and identification system resistant to voltage changes, comprising: a load waveform database construction module, a load waveform feature calculation module, a feature subset filtering module, an identification model construction module, and a real-time load identification module; the load waveform database construction module is used to collect steady-state voltage waveforms and steady-state current waveforms at multiple target voltage points, preprocess the waveforms, and construct a load waveform database for different voltage operating conditions; the load waveform feature calculation module is used to calculate multiple categories of load waveform features from voltage and current waveforms to form feature vectors; the feature subset filtering module is used to group and statistically analyze the feature vectors, and perform joint discrimination based on a multi-statistic integration criterion to obtain feature subsets; the identification model construction module is used to perform data augmentation processing on voltage and current waveforms based on the load waveform database, generate an enhanced sample set, and construct a neural network load identification model; the real-time load identification module is used to calculate real-time feature vectors of the feature subsets based on real-time voltage and current waveforms, input the real-time feature vectors into the neural network load identification model, and output real-time load identification results.

[0017] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method for extracting and identifying load features resistant to voltage changes.

[0018] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for extracting and identifying load features resistant to voltage changes.

[0019] The beneficial effects of this invention are as follows: This invention establishes a load waveform database at multiple target voltage points and performs synchronous acquisition, time alignment, noise reduction filtering, and power frequency cycle segmentation on steady-state voltage and current waveforms. This ensures that the voltage and current waveforms used in subsequent calculations have a consistent time reference, reducing waveform deviations caused by voltage changes, asynchronous sampling, and noise interference. Based on this, this invention calculates multi-category load waveform characteristics based on voltage and current waveforms within the same power frequency cycle and forms feature vectors in a preset order. This allows load characteristics to simultaneously reflect time-domain changes, phase relationships, harmonic distribution, power changes, and waveform shape differences, providing a more complete feature basis for load category discrimination. Furthermore, this invention combines cross-voltage dispersion indices, intra-class consistency indices, and multi-statistic integration criteria to screen features, eliminating features that fluctuate significantly with voltage changes and retaining features that are relatively consistent across different target voltage points, thereby reducing the interference of voltage changes on the model input. Furthermore, this invention generates an enhanced sample set through cross-voltage resampling consistency, noise injection, and phase perturbation, and introduces cross-voltage consistency constraints during model training to keep the output distribution of the same load at different voltage points close, thereby making the neural network load identification model more suitable for actual identification scenarios where the supply voltage changes. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a method for extracting and identifying load features resistant to voltage changes, provided as an embodiment of the present invention.

[0022] Figure 2 The above is a flowchart of a method for extracting and identifying load features resistant to voltage changes, provided as an embodiment of the present invention.

[0023] Figure 3This is a schematic diagram illustrating the calculation of load waveform features and the composition of feature vectors in a load feature extraction and identification method that resists voltage changes, provided as an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the model training process for a load feature extraction and identification method that resists voltage changes, provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0026] Example 1, referring to Figure 1 and Figure 2 This is one embodiment of the present invention, which provides a method for extracting and identifying load features resistant to voltage changes, including: S1. Collect steady-state voltage waveforms and steady-state current waveforms at multiple target voltage points, preprocess the waveforms, and construct a load waveform database under different voltage operating conditions.

[0027] A load waveform database is constructed. The load under test is powered by an adjustable power supply at multiple target voltage points from 198V to 235.4V. The steady-state voltage waveform and steady-state current waveform at each target voltage point are collected synchronously by voltage and current sensors. The collected waveforms are time-aligned, noise-reducing filtered, and divided into power frequency cycles. The processed waveforms are then associated with and stored with the corresponding voltage and load category identifiers to form a load waveform database for different voltage operating conditions.

[0028] S2. Calculate the characteristics of multiple types of load waveforms from voltage and current waveforms to form a feature vector.

[0029] The system calculates the characteristics of diverse load waveforms. For each record in the load waveform database, it calculates the time-domain statistical characteristics, amplitude and phase characteristics, harmonic spectrum characteristics, power and energy characteristics, and waveform shape characteristics based on the voltage and current waveforms within the same power frequency cycle. The system then assembles the characteristics of each category into a feature vector in a preset order.

[0030] S3. Perform group statistics on the feature vectors, and implement joint discrimination based on the multi-statistic integration criterion to obtain feature subsets.

[0031] A subset of features resistant to voltage changes is selected, and the feature vectors are grouped and statistically analyzed according to load category. For each feature, the cross-voltage dispersion index and intra-class consistency index are calculated in the range of 198V~235.4V. Based on the preset multi-statistic integration criterion, the dispersion index and consistency index of each feature are jointly judged, and features whose dispersion with voltage change exceeds the threshold are removed. The feature subset whose variance with voltage change meets the threshold constraint is obtained.

[0032] S4. Based on the load waveform database, perform data augmentation processing on voltage and current waveforms to generate an enhanced sample set and construct a neural network load identification model.

[0033] To train a load identification model resistant to voltage changes, a data augmentation training strategy is introduced. Based on a load waveform database, cross-voltage resampling and consistency, noise injection, and phase perturbation are performed on voltage and current waveforms to generate an augmented sample set. The feature vectors corresponding to the feature subset are used as inputs to the neural network load identification model, and the load category is used as the supervision label for training. During the training process, a loss function with cross-voltage consistency constraints is used to constrain the output distribution of the same load at different voltage points, thus obtaining a neural network load identification model resistant to voltage changes.

[0034] S5. Calculate the real-time feature vector of the feature subset based on the real-time voltage waveform and the real-time current waveform, input the real-time feature vector into the neural network load identification model, and output the real-time load identification result.

[0035] Online load type identification involves collecting real-time voltage and current waveforms from voltage and current sensors in practical applications. The real-time waveforms undergo preprocessing consistent with the database construction steps. Real-time feature vectors corresponding to feature subsets are calculated, and these feature vectors are input into a neural network load identification model resistant to voltage changes, outputting real-time load identification results.

[0036] It should be noted that, as Figure 1 and Figure 2 As shown, this invention does not directly train the load identification model under a single voltage condition. Instead, it first collects steady-state voltage and current waveforms of the same type of load at multiple target voltage points and establishes a load waveform database after unified preprocessing. Therefore, during subsequent feature calculations and model training, the waveform variations of the same load under different supply voltages can be preserved, avoiding the model from only learning waveform features under a fixed voltage.

[0037] Specifically, step S1 is used to form the waveform sample basis under different voltage operating conditions. By performing time alignment, noise filtering, and power frequency cycle segmentation on the acquired waveforms, the voltage and current waveforms used in subsequent calculations can be within the same power frequency cycle, reducing feature deviations caused by asynchronous sampling, noise interference, or inconsistent cycle truncation. Step S2 calculates time-domain statistical characteristics, amplitude and phase characteristics, harmonic spectrum characteristics, power and energy characteristics, and waveform shape characteristics within the same power frequency cycle. This ensures that the feature vectors can reflect not only load amplitude changes but also phase relationships, harmonic distribution, power consumption, and waveform shape differences. Compared to using only the RMS value of current, power, or a single harmonic index, this step provides more comparable feature sources for subsequent screening.

[0038] Step S3 is used to filter features from the aforementioned feature vectors that are less affected by voltage changes. Specifically, the feature vectors are grouped and statistically analyzed according to load category, and a joint judgment is made by combining the cross-voltage dispersion index and the intra-class consistency index. This can eliminate features that fluctuate significantly with voltage changes, while retaining features that still have stable performance at different target voltage points. In this way, during the model training phase, the features input to the model no longer mainly rely on components that are prone to drift with changes in supply voltage, thereby reducing the interference of voltage changes on load category judgment.

[0039] Step S4 further introduces cross-voltage resampling consistency, noise injection, and phase perturbation processing during the model training phase, enabling the training samples to cover situations that may occur in actual operation, such as voltage variations, measurement noise, and phase shifts. Simultaneously, a loss function with cross-voltage consistency constraints constrains the output distribution of the same load at different voltage points, reducing the model's dependence on voltage differences while learning load category differences. Step S5 then uses the aforementioned feature selection and model training results for online identification. By performing consistent preprocessing on real-time voltage and current waveforms and calculating real-time feature vectors for feature subsets, the online input is made consistent with the training input, thus outputting real-time load identification results.

[0040] Therefore, the processing procedure of this invention enables the load identification model to adapt to waveform differences under varying supply voltage. Compared with methods that directly train the model based solely on fixed voltage samples or without filtering features, this invention reduces the impact of feature drift caused by voltage changes on the identification results, making it more suitable for actual power consumption scenarios where supply voltage fluctuates.

[0041] Example 2, refer to Figure 3 and Figure 4 As an embodiment of the present invention, based on the previous embodiment, a method for extracting and identifying load features resistant to voltage changes is provided, including: Furthermore, in step S1, steady-state voltage waveforms and steady-state current waveforms are acquired at multiple target voltage points, the waveforms are preprocessed, and a load waveform database under different voltage operating conditions is constructed, including the following steps A1-A3: A1. The adjustable power supply provides power to the load at multiple target voltage points within the voltage regulation range. The steady-state voltage waveform and steady-state current waveform at each target voltage point are collected synchronously by voltage and current sensors.

[0042] The voltage regulation range refers to 198V~235.4V, which is determined based on the allowable deviation range of the 220V single-phase power supply voltage.

[0043] Specifically, a test circuit is constructed, which includes, in sequence, an adjustable voltage power supply, the load under test, the connection terminals of voltage and current sensors, and a data acquisition device; wherein, the adjustable voltage power supply outputs a single-phase AC fundamental frequency of [frequency value missing]. sinusoidal voltage, Selected during implementation The effective value of the output voltage of the adjustable power supply is within... Within the range, according to the preset voltage step size Set multiple target voltage point sets In implementation Selected as ,correspond In this invention, the parameter values ​​selected in the implementation are the engineering debugging values ​​in this embodiment, and can be adjusted according to the actual application scenario.

[0044] For any target voltage point The power supply and data acquisition process adopts a "step setting - steady-state criterion confirmation - synchronous acquisition" approach. First, the effective value of the output voltage of the adjustable power supply is set to... And remain constant; then enter the steady-state criterion confirmation stage, in continuous The rate of change of the effective voltage and effective current values ​​within each power frequency cycle is calculated, and data acquisition begins only after the steady-state condition is met. In practice... Select 100.

[0045] With sampling frequency Voltage and current are sampled synchronously during implementation. Selected as The number of sample points is based on a preset power frequency cycle. This represents the number of sampling points within one power frequency cycle. satisfy: in, The preset number of sample points for the power frequency cycle, expressed in points, represents the number of points; Sampling frequency, in Hz; This refers to the fundamental frequency of the power supply, measured in Hz. In implementation, and ,correspond point.

[0046] by Represents the power frequency cycle number, defined as the first... RMS voltage value per power frequency cycle With the effective value of current for in, This is a discrete voltage sampling sequence, with units of V; This is a discrete sampling sequence of current, in amperes (A). The power frequency cycle number; The sampling point number within the period; The number of sample points for the preset power frequency cycle.

[0047] In continuous Calculate the rate of change of the effective voltage value within each power frequency cycle. With the rate of change of the effective value of current : in, For the first The rate of change of the effective voltage value per power frequency cycle; For the first Rate of change of effective current value per power frequency cycle; For the first The effective voltage value per power frequency cycle; For the first The effective value of current per power frequency cycle.

[0048] When for all Simultaneously satisfy and When a steady state is reached, data acquisition is initiated at the boundary of the power frequency cycle; during implementation, Select 0.002. The value is selected as 0.005. This steady-state criterion simultaneously constrains both the voltage side and the load response side, avoiding a situation where only the voltage is stable while the load remains in a transient state, thereby ensuring the steady-state consistency of the acquired waveform.

[0049] During the data acquisition phase, voltage and current sensors synchronously acquire data under the same clock on the same data acquisition device, forming a steady-state voltage waveform sequence. With steady-state current waveform sequence To ensure comparability between different target voltage points, the same acquisition duration was set for each target voltage point. and based on the number of sampling points Indicates the data collection length, satisfying... in, This represents the number of sampling points corresponding to the acquisition length, expressed in points. The data collection duration is expressed in seconds (s). This refers to the sampling frequency, measured in Hz. In implementation, Selected as ,correspond Point. Target voltage point. Sampling frequency Power supply fundamental frequency Collection time With sensor range, accuracy and other metadata , Simultaneously, this record provides consistent constraints for subsequent alignment, segmentation, and identifier association.

[0050] A2. Perform time alignment, noise reduction filtering, and power frequency cycle segmentation on the acquired waveform.

[0051] Specifically, the steady-state voltage waveform sequence obtained from A1 With steady-state current waveform sequence The preprocessing flow, with a fixed execution order, includes time alignment, noise reduction filtering, and power frequency cycle segmentation. The output of this flow is a sequence of voltage waveforms with a consistent time base. With current processing waveform sequence and its periodic segment set.

[0052] Time alignment is achieved by maximizing the cross-correlation between voltage and current waveforms to eliminate channel group delays between voltage and current sensors, anti-aliasing filter group delays, and fixed phase shifts caused by wiring. This is done at the integer sampling point level. A cross-correlation function is defined. for: in, Target voltage point The cross-correlation value under the following conditions; Alignment offset, in points; This represents the number of sampling points corresponding to the acquisition length. Target voltage point Next One voltage sampling point; Target voltage point Next One current sampling point; This is the sampling point number.

[0053] Within the preset search range Internal use Maximum As an alignment offset, and according to the current waveform The aligned current waveform is obtained by translating forward: in, Target voltage point The corresponding optimal alignment offset, in points; These are the aligned current sampling points, in amperes (A). To avoid false alignment introduced by over-searching, the implementation is configured as follows: The corresponding maximum alignment range is one power frequency cycle. ,when hour, point.

[0054] Noise reduction filtering is used to suppress the impact of sensor thermal noise and electromagnetic interference on waveform details, while preserving the load current harmonic structure from being disrupted by smoothing. Zero-phase digital bandpass filtering is applied to both the aligned voltage and current waveforms. The cutoff frequency in the passband is set. With passband cutoff frequency And select during implementation , ; Less than the Nyquist frequency To avoid spectrum folding, phase distortion is eliminated through a combination of forward and reverse filtering, resulting in a filtered voltage waveform at output. Compared with the filtered current waveform .

[0055] To ensure that each period segment is built up with the same phase reference, power frequency cycle segmentation employs a "voltage zero-crossing rising edge phase-locked segmentation" strategy, using the voltage waveform as the phase reference. First, the filtered voltage waveform... The test meets the requirements and sampling points This serves as a candidate point for the rising edge of the zero-crossing; subsequently, linear interpolation is used near the candidate point to estimate the fractional sampling offset at the zero-crossing moment. ,in The crossover number is zero. The interpolation calculation is as follows: in, For the first Fractional sampling offset of each zero-crossing rising edge; For the first The candidate sampling point number corresponding to each zero-crossing rising edge; These are the voltage sampling points after filtering; using After subsampling correction of the period start point, the waveform is then... Slicing is performed with each point as a period to form the first... Voltage waveform of each cycle segment With the Current waveform of each cycle segment ;in and The starting points correspond to the same zero-crossing rising edge, ensuring that periodic segments at different target voltage points can be directly compared under the same phase reference. In implementation, continuous... Each periodic segment is used for database construction. Select 100.

[0056] A3. Associate and store the processed waveforms with the corresponding voltage and load category identifiers to form a load waveform database for different voltage operating conditions.

[0057] Specifically, a storage structure is established with "load category identifier, voltage identifier, and power frequency cycle segment sequence number" as the main index, and the set of cycle segments output by A2 is structured and stored in the database. A unique load category identifier is assigned to each measured load. ,in For load category number; The names are provided by a preset load category dictionary and remain consistent throughout the entire library, avoiding different names for the same object. For each target voltage point... Assign a unique voltage identifier and will and The voltage identifier is stored in a one-to-one correspondence to ensure traceability in the database.

[0058] To achieve strong correlation between waveform and identifier storage, a data unit is defined for each inbound record. for: in, To identify by load category With voltage marking The index of One power frequency cycle data unit; This represents the corresponding voltage period segment vector, in units of V; This is the corresponding current period segment vector, in units of Am; The effective value of the voltage at the target voltage point, in V; Sampling frequency, in Hz; This refers to the fundamental frequency of the power supply, measured in Hz. The preset number of sample points for the power frequency cycle; This is the sequence number of the power frequency cycle segment.

[0059] Upon data entry, verification information is written to each data unit to ensure data integrity and tamper-proof consistency. The verification information uses a digest based on waveform values. And stored together with the data unit: in, For the first Summary value of each power frequency cycle data unit; The first voltage period segment Each sampling point is represented by a unit of V. The first of the current periodic segments Each sampling point is represented in units of A. The sampling point number within the period; For voltage summary weighting coefficients; This represents the weighting coefficient for the current summary. and To ensure that the preset constants are consistent across the entire database, the following selections are made during implementation. , By storing This enables consistency verification of the input waveforms during subsequent calls, migrations, and training processes, preventing waveform and identifier mismatches.

[0060] Finally, identify the set by all load categories. With all voltage identifier sets This serves as an index, creating a load waveform database covering different voltage operating conditions; each load category is identified... The following contains all voltage markings. The corresponding waveform records ensure that the steady-state waveforms of the same load category at different target voltage points can be directly queried, compared, and retrieved.

[0061] Furthermore, in step S2, multiple categories of load waveform characteristics are calculated from the voltage and current waveforms to form a feature vector, including the following steps B1-B2: B1. For each record in the load waveform database, calculate the characteristics of multiple categories of load waveforms based on the voltage and current waveforms within the same power frequency cycle.

[0062] Multiple categories of load waveform characteristics include time-domain statistical characteristics, amplitude and phase characteristics, harmonic spectrum characteristics, power and energy characteristics, and waveform shape characteristics. For example... Figure 3 As shown.

[0063] Specifically, for the same power frequency cycle segment, the voltage cycle segment vector is taken separately. With current period segment vector ,in The preset number of sample points for the power frequency cycle; during implementation The value is 240, corresponding to the fundamental frequency of the power supply. Sampling frequency is The following features are all based on the same power frequency cycle. and The calculation ensures the temporal consistency of the features.

[0064] Time-domain statistical characteristics were used to calculate the mean, variance, peak-to-peak value, and gust factor for both voltage and current periodic segments. Taking the voltage periodic segment as an example, its mean... ,variance Peak-to-peak value With peak factor Defined as follows: in, The first voltage period segment One sampling point; The sampling point number within the period; The preset number of sample points for the power frequency cycle; This is the average voltage value; For voltage variance, ; Peak-to-peak voltage; This is the peak voltage factor; This is the effective value of the voltage. The corresponding value on the current side is... , , and ,in For current sampling points, This is the effective value of the current.

[0065] Where the effective value of voltage With the effective value of current They are respectively: in, The first of the current periodic segments One sampling point.

[0066] The amplitude and phase characteristics are obtained by extracting the fundamental frequency amplitude and phase within the same power frequency cycle, using the fundamental frequency of the power supply. Synchronous discrete Fourier projection is used to calculate the fundamental complex coefficients. The fundamental angular frequency is defined. It is the following formula: in, The fundamental angular frequency; Define discrete times within one power frequency cycle, using the fundamental power supply frequency. It is the following formula: in, For the first The time corresponding to each sampling point; The sampling frequency.

[0067] voltage fundamental complex coefficient With the complex coefficient of the fundamental current wave Defined as follows: in, These are the complex coefficients of the fundamental voltage wave; These are the complex coefficients of the fundamental current wave; It is the imaginary unit and satisfies .

[0068] Based on this, the fundamental voltage amplitude is obtained. Current fundamental amplitude and voltage fundamental phase Current fundamental phase : in, This represents the fundamental voltage amplitude. This represents the amplitude of the fundamental current wave. The phase of the voltage fundamental wave; The phase of the fundamental current wave; This is an operation for taking complex phase angles.

[0069] Further calculation of phase difference characteristics : in, The fundamental phase difference is expressed in rad.

[0070] Harmonic spectrum characteristics characterize the nonlinear characteristics of the load under the same target voltage, calculated up to the [missing information]. The amplitude spectrum and total harmonic distortion of the second harmonic. For any harmonic order. Define voltage harmonic complex coefficients With current harmonic complex coefficients It is the following formula: in, The harmonic order is an integer; The largest harmonic order; For the first Second voltage harmonic complex coefficients; For the first The complex coefficients of the second current harmonic; the meanings of the other symbols are the same as above. In implementation, The answer is 25.

[0071] Voltage number Second harmonic amplitude With current Second harmonic amplitude They are respectively: in, For the first Secondary voltage harmonic amplitude; For the first Secondary harmonic amplitude.

[0072] The total harmonic distortion rate is defined using current as an example. It is the following formula: in, The total harmonic distortion of the current; This represents the amplitude of the fundamental current wave. For the first Secondary harmonic amplitude; This is the maximum harmonic order. The voltage side corresponds to... .

[0073] Power and energy characteristics, based on instantaneous power sequences within the same power frequency cycle Calculate active power, apparent power, reactive power, and periodic energy. Instantaneous power. Defined as follows: in, For the first Instantaneous power at each sampling point; For voltage sampling points; This is the current sampling point.

[0074] Average active power during the power frequency cycle It is the following formula: in, This refers to active power.

[0075] Apparent power Defined as follows: in, Apparent power; This is the effective value of the voltage; This is the effective value of the current.

[0076] reactive power The power trigonometric relationship is used to calculate the following formula: in, Reactive power; Apparent power; This refers to active power.

[0077] Power frequency cycle energy Defined as: in, Energy is the power frequency cycle energy; The sampling time interval and .

[0078] Waveform shape characteristics are used to characterize the shape difference and conduction characteristics of the current waveform relative to the voltage waveform. This involves calculating the normalized cross-correlation peak value, waveform asymmetry, and peak density. First, the DC component of the voltage and current is removed to obtain... and : in, These are the voltage sampling points after removing the DC component; These are the current sampling points after removing the DC component; This is the average voltage value; This is the average current.

[0079] Normalized cross-correlation peak Defined as: in, The normalized cross-correlation peak value; This refers to the relevant offset, in points. This represents the maximum relevant offset, expressed in points. In implementation, Selected as ,when hour, Points are used to calculate the waveform shape, and their range is smaller than that used in preprocessing. .

[0080] Waveform asymmetry is defined by the difference between the effective values ​​of the positive and negative half-cycles. Taking current as an example, the effective value of the positive half-cycle is defined as follows: With the effective value of the negative half cycle for: in, To meet The set of sampling point indices; To meet The set of sampling point indices; For set The number of elements; For set The number of elements; This is the effective value of the positive half-cycle of the current; This is the effective value of the negative half-cycle of the current; For current asymmetry; The effective value of the current is given, and the corresponding voltage value is obtained from the voltage side. .

[0081] Peak intensity is defined as the ratio of peak current to RMS current: in, Current spike density; For current sampling points; This is the effective value of the current.

[0082] B2. Assemble the feature vectors from each category of features in a preset order.

[0083] Specifically, define the feature vector. The fixed arrangement order is: time domain statistical feature group, amplitude and phase feature group, harmonic spectrum feature group, power and energy feature group, waveform shape feature group, and the element order and dimension upper limit are defined for each feature group so that the feature vector dimension of different samples output at different target voltage points is completely consistent.

[0084] Similar characteristics from the voltage and current sides are concatenated in the order of "voltage first, current second," and the harmonic spectrum characteristics are arranged from low to high harmonic order, with the highest harmonic order listed first. Define the feature vector as the upper limit. for: in, The feature vector corresponding to a single power frequency cycle sample; For time-domain statistical feature groups; Amplitude and phase characteristic groups; The characteristic group of the harmonic spectrum; Power and energy characteristic groups; This is a waveform shape feature group.

[0085] The feature groups are composed of the following order: in, , , , These are the mean voltage, voltage variance, peak-to-peak voltage, and peak-to-peak voltage factor, respectively. , , , These are the mean current, the variance of current, the peak-to-peak current, and the peak-to-peak current factor, respectively. This is the effective value of the voltage; This is the effective value of the current; This represents the fundamental voltage amplitude. This represents the amplitude of the fundamental current wave. The fundamental phase difference; For the first Secondary voltage harmonic amplitude; For the first Secondary harmonic amplitude; The largest harmonic order; This represents the total harmonic distortion (THD) of the voltage. The total harmonic distortion of the current; Active power; Apparent power; Reactive power; Energy is the power frequency cycle energy; The normalized cross-correlation peak value; Voltage asymmetry; For current asymmetry; This refers to the current spike frequency.

[0086] In implementation, when the maximum harmonic order At that time, the dimension of the harmonic spectrum feature group is dimension.

[0087] Furthermore, in step S3, the feature vectors are grouped and statistically analyzed, and joint discrimination is performed based on the multi-statistic ensemble criterion to obtain feature subsets, including the following steps C1-C2: C1. Group and statistically analyze the feature vectors according to the load category, and calculate the cross-voltage dispersion index and intra-class consistency index for each feature within the voltage regulation range.

[0088] Specifically, let the load category identifier set be... The voltage identifier set is Each voltage is identified With the target voltage RMS value One-to-one correspondence; identifier for any load category Collect all feature vector samples of this load category under each voltage identifier, and denote the first... The feature vector of each sample is ,in, For the feature vector dimension, For load category number, For voltage identification serial number, This is the sample number.

[0089] To perform cross-voltage statistics for each feature dimension, let the feature dimension index be... ,in Definition of the first Dimensional features in load category With voltage marking Sample mean With sample variance for: in, For feature vectors The 3D eigenvalues; For load category Voltage markings The number of samples below; This corresponds to the mean; This represents the corresponding variance.

[0090] The cross-voltage dispersion index is used to characterize the first voltage distribution within the same load category. The amplitude of the characteristic fluctuation with voltage variation. Define load category. The Dimensional characteristic cross-voltage dispersion index for: in, It is an index of cross-voltage dispersion; For the number of voltage indicators; For the first Dimensional features in load category The average value for all voltage markings.

[0091] Intra-class consistency index is used to characterize the first load under the same voltage rating within the same load category. The stability of dimensional features across samples. To ensure the comparability of intra-class consistency indices across different feature dimensions, a voltage-robust scale normalization term is introduced. : in, For the first Dimensional features in load category The following is a robust scale; This is an operation to find the median; To prevent positive constants with a denominator of zero during implementation... Selected as .

[0092] Based on this, the intra-class consistency index is defined. for: in, This is an intra-class consistency indicator, and its value range is... A higher value indicates higher intra-class consistency.

[0093] To facilitate subsequent unified judgment, the cross-voltage dispersion index is normalized according to a robust metric to obtain a normalized cross-voltage dispersion index. : in, This is a normalized cross-voltage dispersion index.

[0094] C2. Based on the preset multi-statistic integration criteria, the dispersion index and consistency index of each feature are jointly judged, and features whose dispersion exceeds the threshold with voltage change are eliminated, so as to obtain a feature subset whose variance with voltage change meets the threshold constraint.

[0095] Specifically, firstly regarding the first Summary statistic of cross-load category dispersion of dimensional features Consistency summary statistics across load categories The calculation is performed and defined as follows: in, For the first The summation statistic of cross-class dispersion of a feature; For the first Cross-class consistency summary statistics for dimensional features; To obtain Quantile operations; To obtain Quantile operations; and This is a preset quantile parameter. In implementation, Selected as 90, Selecting 10 ensures that the screening results remain sensitive to both a few high-discrete load categories and a few low-consistency load categories.

[0096] Secondly, construct integrated discriminant. : in, For the first An integrated discriminant measure of 3D features; and The weighting coefficients are satisfied. ; This is the dispersion threshold; This is the consistency threshold; To prevent positive constants with a denominator of zero, it must be consistent with A1. In implementation, Choose 0.6. Select 0.4; Select 1.0; Choose 0.7.

[0097] The joint discrimination rule is defined as follows: if and only if both conditions are met. and At that time, the judgment of the first The dimensional feature is one whose variance as a function of voltage satisfies a threshold constraint; otherwise, the dimensional feature is determined to be the dimensional feature. Features that exhibit dispersion exceeding a threshold with respect to voltage changes are removed from the feature set.

[0098] Based on this, the feature subset index set is obtained. : in, This is the filtered set of feature indices. Any original feature vector... according to Selected variance-constrained eigenvectors : in, The feature vector after removing highly discrete features; Original feature vector The 1-dimensional eigenvalues.

[0099] Furthermore, in step S4, data augmentation processing is performed on the voltage and current waveforms based on the load waveform database to generate an enhanced sample set and construct a neural network load identification model, including the following steps D1-D3: D1: Introduce a data augmentation training strategy. Based on the load waveform database, perform cross-voltage resampling consistency, noise injection, and phase perturbation data augmentation processing on voltage and current waveforms to generate an augmented sample set.

[0100] Specifically, for any load category identifier in the load waveform database Any voltage marking With any power frequency period segment number Corresponding periodic segment voltage waveform and periodic segment current waveform Within the same power frequency period window, an enhanced operator sequence is constructed.

[0101] Cross-voltage resampling uniformity is used to unify waveforms at different target voltage points to the same equivalent voltage amplitude reference, preventing the network from using voltage amplitude differences as load category criteria. First, the target voltage RMS value is used... RMS value of reference voltage Constructing voltage scaling factor : in, This is the voltage scaling factor; This is the effective value of the reference voltage, in volts (V). Voltage label The corresponding target voltage RMS value, in V. In implementation, Selected as .

[0102] The voltage and current waveforms are scaled by the same factor to achieve amplitude uniformity, resulting in a uniform waveform. and : in, These are the standardized voltage sampling points, in units of V; These are the standardized current sampling points, in amperes (A). The sampling point number within the period and .

[0103] To ensure that the voltage phase reference does not drift after unification, for Perform unit energy normalization and record the normalization coefficients. : in, For uniformity and normalization coefficients; These are the voltage sampling points after standardization and normalization; These are the current sampling points after standardization and normalization. This is achieved by using the same sampling method for both voltage and current. This ensures that the relative amplitude relationship between the two remains unchanged.

[0104] Noise injection is used to simulate sensor noise and electromagnetic interference in metering data acquisition while maintaining controllable noise levels. Zero-mean Gaussian noise is injected into the homogenized voltage and current respectively to obtain noise-enhanced waveforms. and : in, and They are independent and identically distributed standard normal random variables; The standard deviation of voltage noise; This represents the standard deviation of current noise. In implementation, Set as 0.01 , Set as 0.02 The above values ​​are based on the uniform and normalized amplitude dimensions.

[0105] To prevent noise injection from disrupting the steady-state consistency of the single-cycle window, saturation limiting is applied to the noise-enhanced waveform, with saturation thresholds of [missing information]. and : in, It is a saturation constraint function; This is the voltage saturation threshold. This is the current saturation threshold. In implementation, Set as , Set as .

[0106] Set the cyclic shift amount corresponding to the phase perturbation and order From the set of integers Values ​​to be taken from: in, It is a set of cyclic shift values; This represents the maximum displacement, expressed in points. In implementation, Selected as Point, corresponding to The maximum time perturbation is .

[0107] right and Synchronous cyclic shifting yields phase perturbation waveforms and : in, and These are the enhanced voltage sampling points and the enhanced current sampling points, respectively. For modulo operation; This is the cyclic shift amount, in points.

[0108] Following the above process, each original periodic segment sample is generated. One augmented sample is obtained to form an augmented sample set. In implementation, Selected as And retain the original load category identifier for each augmented sample. With voltage marking Used for constructing subsequent consistency constraints.

[0109] D2. Use the feature vectors of the feature subset as input to the neural network load identification model, and use the load category as the supervision label for training.

[0110] Specifically, the enhanced voltage waveform obtained from D1 With enhanced current waveform Calculate the eigenvector within the same power frequency period window and index the set by feature subset Variance-constrained eigenvectors were obtained through screening. : in, To enhance the feature subset vector corresponding to the sample; The feature subset dimension; To enhance the eigenvector of the first 3D eigenvalues; It is a set of feature indices whose variance varies with voltage and satisfies a threshold constraint.

[0111] Establish a neural network load identification model ,in These are the network parameters. The model input is... The model output is a predicted probability distribution vector for all load categories. : in, This is the predicted probability distribution vector; Total number of load categories; For the first The predicted probability of each load category; For load category indexing.

[0112] Identified by load category Corresponding one-hot vector As a supervisory label, cross-entropy loss is used for supervised training, and single-sample classification loss is used. Defined as: in, To enhance the classification loss of the samples; It is the natural logarithm function.

[0113] In implementation, the neural network load identification model A multilayer perceptron structure is adopted, with the number of hidden layers set to . The number of neurons in the hidden layers was set to 128, 64, and 32, respectively. The activation function was ReLU, the optimizer was Adam, and the initial learning rate was set to... The number of training rounds is set to 80.

[0114] D3. During the training process, a loss function with cross-voltage consistency constraints is used to constrain the output distribution of the same load at different voltage points, thereby obtaining a neural network load identification model that is resistant to voltage changes.

[0115] Specifically, for any load category Within the same training batch, enhance samples are sampled from at least two different voltage identifiers to form a cross-voltage sample pair for the same load, and two enhance samples are selected. and ,in The corresponding predicted probability distribution vectors are respectively and .

[0116] Define cross-voltage consistency constraint loss For symmetric KL divergence: in, This is due to the loss of consistency in output distribution for the same load category under different voltage ratings; Total number of load categories; For the sample For the The predicted probability of a class; For the sample For the The predicted probability of a class.

[0117] In calculation When the predicted probability is truncated to a lower bound, the lower bound is 1. : in, The truncated predicted probability; To truncate the lower limit. In implementation, Selected as The truncated probability vector is then normalized again along the category dimension to maintain the sum of probabilities. .

[0118] For all trans-voltage sample pairs within each training batch Take the average value to obtain the batch consistency loss. The total loss is obtained by weighted summation of classification loss and consistency loss. : in, Total loss; This represents the mean of the batch classification loss. This results in batch consistency loss. This is the consistency loss weighting coefficient. In implementation, Set to 0.2.

[0119] Using the total loss Gradient backpropagation updates network parameters The training continues until convergence, resulting in a neural network load identification model that is resistant to voltage changes.

[0120] like Figure 4 As shown, firstly, the voltage and current waveforms in the load waveform database are augmented to generate an augmented sample set; then, a neural network load identification model is trained based on the feature subset vector, and the network parameters are updated by combining classification loss and cross-voltage consistency constraint loss, finally obtaining a neural network load identification model resistant to voltage changes.

[0121] Furthermore, in step S5, the real-time feature vector of the feature subset is calculated based on the real-time voltage waveform and the real-time current waveform. The real-time feature vector is then input into the neural network load identification model, and the real-time load identification result is output, including the following steps E1-E4: E1. In practical application scenarios, voltage and current sensors are used to collect real-time voltage and current waveforms.

[0122] Specifically, voltage and current sensors are deployed at the user's connection point with the power grid, and the outputs of both sensors are connected to the synchronous sampling channel of the same data acquisition device. Real-time voltage and current waveforms are then synchronously acquired using the same sampling clock. Let the real-time sampling frequency be... In implementation Selected as Let the fundamental frequency of the power supply be... In implementation Selected as .

[0123] The real-time voltage discrete sampling sequence is denoted as... The real-time current discrete sampling sequence is denoted as ,in This is the real-time sampling point number. To ensure stable segmentation according to the power frequency cycle, real-time sampling is performed continuously and buffered without interruption. The buffer length is set to [value missing]. Points, in implementation Selected as point.

[0124] E2. Perform preprocessing on the real-time waveform in the same way as the database construction steps.

[0125] Specifically, for the real-time voltage discrete sampling sequence With real-time current discrete sampling sequence The same preprocessing procedure as the load waveform database construction phase is executed, and the order of the preprocessing procedure is fixed as follows: time alignment, noise reduction filtering, and power frequency cycle segmentation.

[0126] Time alignment within the preset search range Internal calculation of real-time cross-correlation function : in, This is the real-time cross-correlation value; For alignment offset; This is the maximum alignment offset, in points. This represents the number of sampling points corresponding to the real-time cache length. The sampling point number within the cache; For the first One real-time voltage sampling point; For the first One real-time current sampling point.

[0127] Take Maximum As an alignment offset, and by adjusting the real-time current sequence according to... The aligned real-time current sequence is obtained by forward translation. In implementation, Set as ,when hour point.

[0128] Denoising filtering for real-time voltage sequences Aligned real-time current sequence Zero-phase digital bandpass filtering is used, and the cutoff frequency in the passband is set to [value missing]. The passband cutoff frequency is set to In implementation , The filtered real-time voltage sequence is obtained. With the filtered real-time current sequence .

[0129] Power frequency cycle segmentation, with As a phase reference, candidate points for zero-crossing rising edges are detected, and linear interpolation is used to estimate the fractional sampling offset of the zero-crossing rising edge. : in, For the first Fractional sampling offset of a real-time zero-crossing rising edge; For the first The candidate sampling point number corresponding to each real-time zero-crossing rising edge; Real-time zero-crossing sequence number; These are the real-time voltage sampling points after filtering.

[0130] Based on the zero-crossing rising edge, the real-time waveform is divided into cycles. Slice the points to obtain the first... Real-time periodic segment voltage waveform With the Real-time periodic segment current waveform The two origins correspond to the same zero-crossing rising edge.

[0131] E3. Calculate the real-time feature vectors corresponding to the feature subsets.

[0132] Specifically, for any real-time periodic segment and Real-time feature vectors are obtained within the same power frequency cycle window using a consistent feature calculation process from the training phase. Then index the set by feature subset. right The real-time feature subset vector is obtained by filtering. : in, A real-time feature subset vector; The feature subset dimension; For the real-time feature vector, the first 3D eigenvalues; It is a set of feature indices whose variance varies with voltage and satisfies a threshold constraint.

[0133] To ensure that the real-time input distribution is consistent with the training input distribution, Perform feature standardization, using the mean vector obtained from the training set as the standardization parameter. with standard deviation vector : in, For the standardized first 3D real-time feature values; Indexing the feature subset dimension and ; For the training set The mean of the dimensional features; For the training set Standard deviation of dimensional features; To prevent positive constants with a denominator of zero during implementation... Selected as .

[0134] E4. Input the real-time feature vector into the neural network load identification model that is resistant to voltage changes, and output the real-time load identification result.

[0135] Specifically, the standardized real-time feature subset vector A neural network load identification model that is resistant to voltage variations To obtain the real-time predicted probability distribution vector : in, This is the real-time predicted probability distribution vector; Total number of load categories; These are the network parameters obtained during training; For the first Predicted probabilities for each load category.

[0136] The load category index corresponding to the highest probability is used as the real-time identification result. : in, For the first Load identification results for each real-time periodic segment; This is an operation that retrieves the index corresponding to the maximum value.

[0137] For continuous The identification results of each real-time periodic segment are subject to majority voting to output the final real-time load identification result. : in, This is the final real-time load identification result; The length of the voting window; This is a calculation for finding the mode; in practice, Selected as ,correspond The output stable window.

[0138] Example 3 is an embodiment of the present invention. This embodiment provides a load feature extraction and identification system resistant to voltage changes, including a load waveform database construction module, a load waveform feature calculation module, a feature subset screening module, an identification model construction module, and a real-time load identification module.

[0139] The load waveform database construction module is used to collect steady-state voltage waveforms and steady-state current waveforms at multiple target voltage points, preprocess the waveforms, and construct a load waveform database under different voltage operating conditions.

[0140] The load waveform feature calculation module is used to calculate multiple categories of load waveform features from voltage and current waveforms, forming a feature vector.

[0141] The feature subset selection module is used to group and statistically analyze feature vectors, and perform joint discrimination based on the multi-statistic integration criterion to obtain feature subsets.

[0142] The identification model building module is used to perform data augmentation processing on voltage and current waveforms based on the load waveform database, generate an enhanced sample set, and build a neural network load identification model.

[0143] The real-time load identification module is used to calculate the real-time feature vector of the feature subset based on the real-time voltage waveform and the real-time current waveform, input the real-time feature vector into the neural network load identification model, and output the real-time load identification result.

[0144] This embodiment also provides an electronic device applicable to a load feature extraction and identification method resistant to voltage changes, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the load feature extraction and identification method resistant to voltage changes proposed in the above embodiment.

[0145] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a load feature extraction and identification method resistant to voltage changes as proposed in the above embodiments.

[0146] The storage medium proposed in this embodiment and the method for extracting and identifying load features that resist voltage changes proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0147] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0148] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for extracting and identifying load characteristics resistant to voltage changes, characterized in that, include: Steady-state voltage and current waveforms were collected at multiple target voltage points, and the waveforms were preprocessed to construct a load waveform database under different voltage operating conditions. Multiple categories of load waveform characteristics are calculated from voltage and current waveforms to form a feature vector; The feature vectors are grouped and statistically analyzed, and joint discrimination is performed based on the multi-statistic ensemble criterion to obtain feature subsets; Based on the load waveform database, voltage and current waveforms are augmented to generate an enhanced sample set and a neural network load identification model is constructed. The real-time feature vectors of the feature subsets are calculated based on the real-time voltage and current waveforms. The real-time feature vectors are then input into the neural network load identification model, which outputs the real-time load identification results.

2. The method for extracting and identifying load characteristics resistant to voltage changes as described in claim 1, characterized in that, The process of performing data augmentation on voltage and current waveforms based on a load waveform database to generate an enhanced sample set and constructing a neural network load identification model includes: A data augmentation training strategy is introduced, which performs cross-voltage resampling consistency, noise injection and phase perturbation data augmentation processing on voltage and current waveforms based on the load waveform database to generate an augmented sample set. The feature vectors of the feature subset are used as input to the neural network load identification model, and the load category is used as the supervision label for training. During the training process, a loss function with cross-voltage consistency constraints is used to constrain the output distribution of the same load at different voltage points, thereby obtaining a neural network load identification model that is resistant to voltage changes.

3. The method for extracting and identifying load characteristics resistant to voltage changes as described in claim 2, characterized in that, The process involves acquiring steady-state voltage and current waveforms at multiple target voltage points, preprocessing the waveforms, and constructing a load waveform database under different voltage operating conditions, including: The load is supplied with power according to the multiple target voltage points within the voltage regulation range of the adjustable power supply. The steady-state voltage waveform and steady-state current waveform at each target voltage point are collected synchronously by voltage sensor and current sensor. The acquired waveforms are processed by time alignment, noise reduction filtering, and power frequency cycle segmentation. The processed waveforms are associated with and stored with the corresponding voltage and load category identifiers to form a load waveform database for different voltage operating conditions.

4. The method for extracting and identifying load characteristics resistant to voltage changes as described in claim 3, characterized in that, The calculation of multiple categories of load waveform characteristics based on voltage and current waveforms The feature vector consists of: For each record in the load waveform database, multiple categories of load waveform characteristics are calculated based on the voltage and current waveforms within the same power frequency cycle. The features of each category are arranged into a feature vector in a preset order.

5. The method for extracting and identifying load characteristics resistant to voltage changes as described in claim 4, characterized in that, The process of grouping and statistically analyzing feature vectors, and then performing joint discrimination based on a multi-statistic ensemble criterion, yields feature subsets, including: The feature vectors are grouped and statistically analyzed according to load category. For each feature, the cross-voltage dispersion index and intra-class consistency index are calculated within the voltage regulation range. Based on the preset multi-statistic integration criteria, the dispersion index and consistency index of each feature are jointly judged, and features whose dispersion exceeds the threshold with voltage change are eliminated, resulting in a feature subset whose variance with voltage change meets the threshold constraint.

6. The method for extracting and identifying load characteristics resistant to voltage changes as described in claim 5, characterized in that, The process of calculating a real-time feature vector from a feature subset based on real-time voltage and current waveforms, inputting the real-time feature vector into a neural network load identification model, and outputting a real-time load identification result includes: Real-time voltage and current waveforms are acquired by voltage and current sensors. Perform preprocessing on the real-time waveform; Calculate the real-time feature vector of the feature subset; The real-time feature vector is input into the neural network load identification model that is resistant to voltage changes, and the real-time load identification result is output.

7. The method for extracting and identifying load characteristics resistant to voltage changes as described in claim 6, characterized in that, For each record in the load waveform database, the calculation of multiple categories of load waveform characteristics based on the voltage and current waveforms within the same power frequency cycle includes: Extract voltage periodic segment vectors and current periodic segment vectors from the same power frequency cycle; The calculation of various load waveform characteristics includes time-domain statistical characteristics, amplitude and phase characteristics, harmonic spectrum characteristics, power and energy characteristics, and waveform shape characteristics; Time-domain statistical features are calculated based on voltage periodic segment vectors and current periodic segment vectors. These time-domain statistical features include mean, variance, peak-to-peak value, and peak factor. The amplitude and phase characteristics are calculated based on the discrete Fourier projection synchronized with the fundamental frequency of the power supply. The amplitude and phase characteristics include the fundamental voltage amplitude, fundamental current amplitude, fundamental voltage phase, fundamental current phase, and fundamental phase difference. The harmonic spectrum characteristics are calculated based on the preset harmonic order. The harmonic spectrum characteristics include voltage harmonic amplitude, current harmonic amplitude, total voltage harmonic distortion rate, and total current harmonic distortion rate. The power and energy characteristics are calculated based on the instantaneous power sequence, and the power and energy characteristics include active power, apparent power, reactive power and power frequency cycle energy. The waveform shape characteristics are calculated based on the voltage and current sampling points after removing the DC component. The waveform shape characteristics include the normalized cross-correlation peak value, waveform asymmetry, and peak value.

8. A load feature extraction and identification system resistant to voltage changes, employing the load feature extraction and identification method resistant to voltage changes as described in any one of claims 1 to 7, characterized in that, include: The system includes a load waveform database construction module, a load waveform feature calculation module, a feature subset filtering module, an identification model construction module, and a real-time load identification module. The load waveform database construction module is used to collect steady-state voltage waveforms and steady-state current waveforms at multiple target voltage points, preprocess the waveforms, and construct a load waveform database under different voltage operating conditions. The load waveform characteristic calculation module is used to calculate multiple categories of load waveform characteristics from voltage and current waveforms. Composition of feature vectors; The feature subset filtering module is used to group and statistically analyze the feature vectors, and perform joint discrimination based on the multi-statistic integration criterion to obtain feature subsets; The identification model construction module is used to perform data augmentation processing on voltage and current waveforms based on the load waveform database, generate an enhanced sample set, and construct a neural network load identification model. The real-time load identification module is used to calculate the real-time feature vector of the feature subset based on the real-time voltage waveform and the real-time current waveform, input the real-time feature vector into the neural network load identification model, and output the real-time load identification result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the load feature extraction and identification method for resisting voltage changes according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the load feature extraction and identification method for resisting voltage changes according to any one of claims 1 to 7.

Images