Nonlinear load identification method and system based on time-frequency analysis

Abstract

The invention belongs to the field of non-intrusive power load monitoring, and discloses a non-intrusive power load identification method, which comprises the following steps of: constructing a load identification framework utilizing frequency domain characteristics, and aims to solve the problem that transient characteristics cannot be effectively applied in non-linear load identification in non-intrusive load monitoring. A current feature extraction technology based on fast Fourier transform and Hilbert-Huang transform and a classification method based on an extreme learning machine as a main body are used for training and testing a public data set to verify that the method is used for extracting transient load features and identifying nonlinear loads.

Description

Technical Field

[0001] This invention belongs to the field of non-intrusive power load monitoring, and specifically relates to a nonlinear load identification method and system based on time-frequency analysis. Background Technology

[0002] Electricity has become the most important end-use energy source in modern society. To address the global energy shortage, achieve social decarbonization, and improve the reliability of power supply and the efficiency of electricity utilization on the user side, research and practice on smart grids are being conducted worldwide. On the demand side of the grid, the widespread application of smart electronic devices and high-energy-consuming equipment has significantly increased the proportion of nonlinear loads at the grid's demand end. How to obtain the load characteristics of nonlinear loads from total user energy consumption data and accurately identify them is a key focus for the power sector.

[0003] Non-intrusive Load Monitoring (NILM) technology requires only one sensor with communication capabilities installed at the power supply inlet of the electrical load. By analyzing the total load power consumption data, it can obtain power consumption information for each (major) appliance within the user's premises. NILM technology is of great significance for improving the observability of demand-side electricity. Its deployment can contribute to the construction of a clean, low-carbon, safe, and efficient energy system and the achievement of "dual carbon" goals, promote pollution and carbon reduction in key industries, and expand and deepen the application of electricity data in social services.

[0004] Since its inception in the late 20th century, the Nonlinear Load Identification (NILM) technique has received widespread attention. Currently, various methods within this field have been applied to nonlinear load identification, with time-frequency domain analysis theory becoming a crucial tool for NILM technology in achieving nonlinear load identification. The Fast Fourier Transform (FFT) and Hilbert-Huang Transform (HFT) theories, matured in speech signal processing, extract target load features from high-frequency load signals to support subsequent load identification. However, due to the diverse types of user-side loads, the widespread occurrence of load aliasing in complex scenarios, and the significant interference of noise, accurately acquiring and identifying the characteristics of the target nonlinear load is challenging. To address these issues, this invention improves upon time-frequency analysis theory, proposing a complete nonlinear load identification architecture based on an Extreme Learning Machine (ELM) model. The classification model is trained and tested using publicly available datasets. Results show that the method of this invention can accurately identify complex nonlinear loads on the user side, achieving a leading position in the field. Summary of the Invention

[0005] In view of the aforementioned existing problems, this invention is proposed. This invention proposes a high-frequency load feature extraction technique based on time-frequency analysis, and simultaneously proposes a nonlinear load identification method based on extreme learning machine.

[0006] To address the aforementioned technical problems, a nonlinear load identification method based on time-frequency analysis is proposed, including:

[0007] Data is collected and preprocessed. The target time series is extracted using a sliding window through a multi-transient load feature extraction module. Time-frequency domain features are extracted based on Fast Fourier Transform and Hilbert-Huang Transform, respectively. Based on the extracted time-domain and frequency-domain multi-transient load features, a multi-feature fusion classification learning model is established. The nonlinear load identification is trained and adjusted using a public dataset. The sliding window length is adjusted based on the effect of the extracted time-frequency features and the load identification effect to obtain the optimal test results and optimize the model in combination with load characteristics.

[0008] As a preferred embodiment of the nonlinear load identification method based on time-frequency analysis described in this invention, the data acquisition and preprocessing include collecting data through current sensors, voltage sensors, temperature sensors, and data acquisition devices; preprocessing the raw data includes denoising, normalization, and synchronization; extracting features from different modal data and encoding them into high-dimensional vectors; and comprehensively analyzing load characteristics under different modalities through a multimodal fusion strategy based on an attention mechanism: the feature vectors of different modalities are input into an attention layer, the attention layer calculates the correlation between different features through a weight matrix to obtain a weight vector, the weight vector is weighted and summed with the feature vectors of each modality to obtain a fused feature vector, and the fused feature vector is input into a learning model for nonlinear load identification.

[0009] As a preferred embodiment of the nonlinear load identification method based on time-frequency analysis described in this invention, the extraction of time-frequency domain features includes truncating the target time series by a sliding window length and performing feature extraction on the truncated series. The feature extraction includes time-domain feature extraction and frequency-domain feature extraction to obtain complete feature samples to the greatest extent.

[0010] The initial sliding window length is defined as N. Based on this window length, the total current signal of the user is truncated, resulting in the target current time series as follows:

[0011] {i(n)|n=1,2,…,N-1,N}

[0012] The frequency domain feature extraction includes extracting frequency domain features, including frequency domain amplitude FA and frequency domain phase FP, from the target current time series using short-time Fourier transform. The original current signal is processed by short-time Fourier transform as follows:

[0013]

[0014] The frequency domain amplitude FA and frequency domain phase FP are calculated as follows:

[0015]

[0016] FR(m)=arctan(ImF(m) / ReF(m))

[0017] Where i(n) is the instantaneous current value at each time point n of the current time series sample, F(m) is the frequency domain signal of the original current signal i(n) of the user's total port, Re(.) and Im(.) are the real part and imaginary part of the target, respectively, m represents the harmonic order, and j is the imaginary unit.

[0018] As a preferred embodiment of the nonlinear load identification method based on time-frequency analysis described in this invention, the time-domain feature extraction includes processing the acquired user port raw current signal i(n) using Hilbert-Huang transform, and the time-domain features include instantaneous amplitude IA, instantaneous phase IP, and instantaneous frequency IF.

[0019] The Hilbert-Huang transform processing of the original current signal i(n) at the main port is as follows:

[0020]

[0021] After Hilbert-Huang transform processing, the original current signal i(t) is analyzed and characterized, and its signal form is analyzed as follows:

[0022]

[0023] The two time-domain features extracted, instantaneous amplitude IA and instantaneous phase IP, are shown below:

[0024]

[0025] IP(t) = arctan(HT(i(t)) / i(t))

[0026] After obtaining the instantaneous amplitude IA and instantaneous phase IP, the following calculations are performed:

[0027]

[0028] Where j represents the imaginary unit, This indicates that the first derivative is obtained in time. Since the obtained characteristic instantaneous phase IP is a discrete time series in form, the derivative result is the difference between two data points, and u represents the integrand variable.

[0029] As a preferred embodiment of the nonlinear load identification method based on time-frequency analysis described in this invention, the classification learning model includes an extreme learning machine model as the main body, which completes the identification of the load by fusing multiple features of the learning input. The extreme learning machine consists of an output layer, an input layer and multiple hidden layers. In the structure of the network, the weight values ​​and bias values ​​between the input layer and the hidden layer are randomly generated and are not iteratively modified.

[0030] Output value t of the output layer node k With input layer node value x i The mathematical relationship is as follows:

[0031]

[0032] When ω is determined j and b j Next, calculate the output matrix H of the hidden layer:

[0033]

[0034] The time-frequency domain features are fused to complete the subsequent model training. A fully connected network is adopted, and the results are used to train the established extreme learning machine model.

[0035] Where l represents the hidden layer number in the network, ω l and b l These are the weight matrix and bias vector from the input layer to the l-th hidden layer node in the network structure, respectively. g(·) refers to the activation function, and β... l This refers to the weight values ​​from the l-th hidden layer node to the output layer node, where L is the number of hidden layer nodes and T is the output matrix of the output layer. Let H be the pseudo-inverse matrix.

[0036] As a preferred embodiment of the nonlinear load identification method based on time-frequency analysis described in this invention, the training and adjustment of the nonlinear load identification includes continuously adjusting the sliding window length based on the time-frequency features extracted at each window length and the load identification effect, using the 5-Fold algorithm for cross-validation, and adjusting the weighting parameters R of the feature fusion connection network.

[0037]

[0038] When E avg (W new )≥E and E avg (W new )-E avg (W old When )≥ΔE, accept the new weighting parameter, when Eavg (W new )<E or E avg (W new )-E avg (W old ) When <ΔE, reject the new weighted parameter;

[0039] Randomly and uniformly divide the original subset into five subsets, select one subset as the test set, and the remaining four subsets as the training set. After training the above model, evaluate it on the test set, and use accuracy and precision as performance evaluation metrics. Repeat the training and testing of the model five times, using each subset as the test set in turn for inspection, ensuring that each subset is used as the test set once for model evaluation. Take the average value of the performance metric results as the final result. Adjust the weight parameter to make the performance metric result after cross-validation the best and meet the requirement of the corresponding target accuracy τ. When the cross-validation ends, verify whether the target accuracy value is reached. If the target accuracy value is reached, output the final result and end the training. If the target accuracy value is not reached, adjust the parameters and perform feature fusion and training again until the target accuracy value is reached and the training ends.

[0040] Among them, W opt is the weighted parameter optimized by minimizing the average loss and the exponential decay of precision. X i represents the i-th subset in X. L(F(x; W), y) measures the difference between the prediction F(x; W) and the true label y. W new is the adjusted weighted parameter. W old is the original weighted parameter. E avg is the average performance metric, ΔE is the performance improvement threshold, and E is the set performance threshold.

[0041] As a preferred scheme of the non-linear load identification method based on time-frequency analysis described in the present invention, wherein: the optimization by combining load characteristics includes that the model automatically adjusts its internal structure and parameters according to the time-frequency characteristics of the load to ensure that the identification effect is always in the optimal state, monitors and compensates for the influence of environmental factors on load characteristics;

[0042]

[0043] Among them, a represents the threshold of the load characteristic extraction accuracy. When it is lower than the threshold, trigger the retraining process of the model. L(t, f) represents the time-frequency characteristic function of the load. H is used to adjust the filter parameters according to the time-frequency characteristics of the load. G is used to compensate for the influence of environmental factors on load characteristics. μ, σ 2 respectively represent the mean and variance of load characteristics.

[0044] When the value of L(t,f) is large, it indicates that the time-frequency characteristics of the load are consistent with the model prediction, and the identification accuracy is high. When the value of L(t,f) is small, it indicates that the time-frequency characteristics of the load differ significantly from the model prediction, and the identification accuracy is low, requiring the triggering of an adaptive update process. When the time-frequency characteristics of the load are abnormal, time-frequency feature analysis is performed, and the filter parameters in the feature extraction module are automatically adjusted to enhance the extraction of frequency band features. When the load characteristics change significantly over time, the adaptive update of the classification learning model is automatically triggered to maintain the identification accuracy. The optimized model is then reapplied to the load identification task to continue monitoring changes in load characteristics and environmental factors. When the adaptive update conditions are met, the next round of optimization process begins.

[0045] The adaptive update conditions include setting a retraining process for the model when the accuracy of load feature extraction is lower than 'a', so as to ensure that the model can adapt to the new load characteristics, and updating the model every 24 hours to improve the model's recognition.

[0046] Another objective of this invention is to provide a nonlinear load identification system based on time-frequency analysis. This invention creates an efficient, adaptive, and accurate nonlinear load identification system that can maintain high identification accuracy and stability under different environmental and load characteristic changes, providing support for the monitoring, control, and protection of power systems.

[0047] As a preferred embodiment of the nonlinear load identification system based on time-frequency analysis described in this invention, it is characterized by comprising a data acquisition module, a multi-transient load feature extraction module, an extreme learning machine network construction module, a model training and adjustment module, and a performance evaluation module.

[0048] The data acquisition module collects current time series data from the user's main outlet to obtain the raw current signal of the nonlinear load, providing data support for subsequent feature extraction and model training.

[0049] The multi-transient load feature extraction module extracts time-frequency features from the collected current time series data, including frequency domain amplitude, frequency domain phase, instantaneous amplitude, instantaneous phase, and instantaneous frequency, transforming the original current signal into an easily identifiable feature vector and improving the accuracy of load identification.

[0050] The Extreme Learning Machine (ELM) network construction module constructs an ELM network containing an input layer, a hidden layer, and an output layer, effectively fuses the multi-time-frequency domain features output by the feature extraction module, and begins subsequent model training.

[0051] The model training and adjustment module trains the Extreme Learning Machine network, adjusts the network weights and biases to achieve good performance on the training data, performs 5-Fold cross-validation on the model, uses each subset as a test set for model evaluation, and adjusts the model based on the results.

[0052] The performance evaluation module evaluates the performance of the trained model on the test set, quantifies the performance of the entire load identification system, and provides a basis for model optimization and adjustment.

[0053] A computer device includes a memory and a processor, the memory storing a computer program, characterized in that the processor executes the computer program to implement the steps of a method for identifying nonlinear loads based on time-frequency analysis.

[0054] A computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the method for identifying nonlinear loads based on time-frequency analysis.

[0055] The beneficial effects of this invention are as follows: By extracting and fusing multi-dimensional frequency domain features, this invention enhances the identification of different load devices; based on this data, by constructing and training an extreme learning model, it achieves accurate identification of the operating status of different electrical appliances, improves the NILM system's ability to perceive the status of electrical appliances in the user's environment, and provides data support for subsequent customized user-specific energy-saving and emission-reduction plans as well as load response plans.

[0056] This invention presents a nonlinear load identification method based on time-frequency analysis. By collecting and preprocessing raw data, and utilizing a multi-transient load feature extraction module and time-frequency domain feature extraction, it can more comprehensively and accurately identify nonlinear loads. This method establishes a multi-feature fusion classification learning model by fusing multi-transient load features in the time and frequency domains, improving the accuracy and reliability of load identification. Simultaneously, the model can automatically adjust its internal structure and parameters based on the time-frequency characteristics of the load, ensuring that the identification effect is always optimal, and monitors and compensates for the impact of environmental factors on load characteristics, improving the model's robustness. By adjusting the sliding window length to obtain optimal test results and optimizing in conjunction with load characteristics, the model can accurately identify nonlinear loads in real time. Furthermore, the 5-Fold algorithm is used for cross-validation, and the weighting parameters of the feature fusion connection network are adjusted, improving the model's training efficiency and performance. When the load feature extraction accuracy falls below a set threshold, the model is retrained to adapt to new load characteristics. The model is updated every 24 hours, improving its recognizability. The multimodal fusion strategy combined with an attention mechanism can comprehensively analyze load characteristics under different modalities, improving the model's generalization ability. The required data acquisition equipment is simple, making it easy to promote in practical applications, reducing operation and maintenance costs, improving power grid stability, and providing strong support for power system operation and management. Attached Figure Description

[0057] 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, wherein:

[0058] Figure 1 The above is a flowchart of a nonlinear load identification method based on time-frequency analysis, provided as an embodiment of the present invention.

[0059] Figure 2 This invention provides an extreme learning machine network structure for a nonlinear load identification method based on time-frequency analysis, as an embodiment of the present invention.

[0060] Figure 3 This is a system functional architecture diagram of a nonlinear load identification system based on time-frequency analysis, provided as an embodiment of the present invention. Detailed Implementation

[0061] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, 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, and 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.

[0062] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0063] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it an embodiment that is mutually exclusive, either alone or selectively, with other embodiments.

[0064] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0065] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0066] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0067] Example 1

[0068] Reference Figures 1-2This is the first embodiment of the present invention, which provides a nonlinear load identification method based on time-frequency analysis, including:

[0069] S1: Collect data and preprocess it. The target time series is extracted by sliding window through the multi-transient load feature extraction module. Time-frequency domain features are extracted based on fast Fourier transform and Hilbert-Huang transform respectively.

[0070] It should be noted that data is collected through current sensors, voltage sensors, temperature sensors, and data acquisition equipment. The data acquisition equipment converts analog signals into digital signals and transmits them to the preprocessing system. The digital signals are preprocessed, including denoising, normalization, and synchronization. Features are extracted from different modal data and encoded into high-dimensional vectors, which are used as input to the machine learning model. A multimodal fusion strategy based on an attention mechanism is used to comprehensively analyze the load characteristics under different modalities: the feature vectors of different modalities are input into the attention layer, which calculates the correlation between different features through a weight matrix to obtain a weight vector. The weight vector is then weighted and summed with the feature vectors of each modality to obtain the fused feature vector. The fused feature vector is then input into the learning model for the identification of nonlinear loads.

[0071] Furthermore, the target time series is truncated by the sliding window length, and features are extracted from the truncated sequence. Feature extraction includes time domain feature extraction and frequency domain feature extraction to obtain complete feature samples to the greatest extent.

[0072] The initial sliding window length is defined as N. Based on this window length, the total current signal of the user is truncated, resulting in the target current time series as follows:

[0073] {i(n)|n=1,2,...,N-1,N}

[0074] For the current time series given in the above formula, the short-time Fourier transform, a variant of the discrete Fourier transform commonly used in the field, is used to extract frequency domain features from the target current signal.

[0075] The frequency domain feature extraction includes extracting frequency domain features, including frequency domain amplitude FA and frequency domain phase FP, from the target current time series using short-time Fourier transform. The original current signal is processed by short-time Fourier transform as follows:

[0076]

[0077] The frequency domain amplitude FA and frequency domain phase FP are calculated as follows:

[0078]

[0079] FR(m)=arctan(ImF(m) / ReF(m))

[0080] Where i(n) is the instantaneous current value at each time point n of the current time series sample, F(m) is the frequency domain signal of the original current signal i(n) of the user's total port, Re(.) and Im(.) are the real part and imaginary part of the target, respectively, m represents the harmonic order, and j is the imaginary unit.

[0081] It should also be noted that, for the extraction of time-domain features, the Hilbert-Huang transform is used to process the acquired user port raw current signal i(n), and the time-domain features include instantaneous amplitude IA, instantaneous phase IP, and instantaneous frequency IF.

[0082] The Hilbert-Huang transform processing of the original current signal i(n) at the main port is as follows:

[0083]

[0084] After Hilbert-Huang transform processing, the original current signal i(t) is analyzed and characterized, and its signal form is analyzed as follows:

[0085]

[0086] The two time-domain features extracted, instantaneous amplitude IA and instantaneous phase IP, are shown below:

[0087]

[0088] IP(t) = arctan(HT(i(t)) / i(t))

[0089] After obtaining the instantaneous amplitude IA and instantaneous phase IP, the following calculations are performed:

[0090]

[0091] Where j represents the imaginary unit, This indicates that the first derivative is obtained in time. Since the obtained characteristic instantaneous phase IP is a discrete time series in form, the derivative result is the difference between two data points, and u represents the integrand variable.

[0092] S2: Based on the extracted time-domain and frequency-domain multi-transient load features, establish a multi-feature fusion classification learning model.

[0093] Furthermore, using an Extreme Learning Machine (ELM) model as the main component, the system identifies the load by fusing multiple features from the learning input. The ELM consists of an output layer, an input layer, and multiple hidden layers. In the network structure, the weights and biases between the input layer and the hidden layers are randomly generated and not iteratively modified. After establishing the learning model, feature fusion is performed based on the multiple time-frequency domain features of the output, and subsequent model training begins.

[0094] Output value t of output layer node k With input layer node value x i The mathematical relationship is as follows:

[0095]

[0096] When ω is determined j and b j Next, the output matrix H of the hidden layer is calculated:

[0097]

[0098] The time-frequency domain features are fused to complete the subsequent model training. A fully connected network is adopted, and the results are used to train the established extreme learning machine model.

[0099] Where l represents the hidden layer number in the network, ω l and b l These are the weight matrix and bias vector from the input layer to the l-th hidden layer node in the network structure, respectively. g(·) refers to the activation function, and β... l This refers to the weight values ​​from the l-th hidden layer node to the output layer node, where L is the number of hidden layer nodes and T is the output matrix of the output layer. Let H be the pseudo-inverse matrix.

[0100] S3: Use public datasets to train and adjust nonlinear load identification, adjust the sliding window length based on the extracted time-frequency feature effect and load identification effect to obtain the optimal test effect and optimize it in combination with load characteristics.

[0101] Furthermore, the sliding window length is continuously adjusted based on the extracted time-frequency features and load identification performance at each window length. A 5-Fold algorithm is used for cross-validation, and the weighting parameters R of the feature fusion connection network are adjusted.

[0102]

[0103] When E avg (W new )≥E and E avg (W new )-Eavg (W old ) ≥ ΔE-time, accept the new weighted parameter, when, E avg (W new ) < E or E avg (W new ) - E avg (W old ) < ΔE-time, reject the new weighted parameter.

[0104] Randomly and evenly divide the original subset into five subsets, select one subset as the test set, and the remaining four subsets as the training set. After training the above model, evaluate it on the test set, use accuracy and precision as performance evaluation metrics, repeat the training and testing of the model five times, use each subset as the test set in turn for inspection, ensure that each subset is used as the test set for model evaluation once, take the average value of the performance metric results as the final result, adjust the weighted parameters of the connection network for feature fusion to improve the overall load recognition effect, preset the target accuracy τ, if the accuracy of the current model is less than the target accuracy, adjust the network parameters until the accuracy reaches the target.

[0105] Among them, W opt is the weighted parameter optimized by minimizing the average loss and exponential decay of precision, X i represents the i-th subset in X, L(F(x; W), y) measures the difference between the prediction F(x; W) and the true label y, W new is the adjusted weighted parameter, W old is the original weighted parameter, E avg is the average performance metric, ΔE is the performance improvement threshold, and E is the set performance threshold.

[0106] It should be noted that the model will automatically adjust its internal structure and parameters according to the time-frequency characteristics of the load to ensure that the recognition effect is always in the optimal state, monitor and compensate for the influence of environmental factors on the load characteristics;

[0107]

[0108] Among them, a represents the threshold of the load feature extraction accuracy. When it is lower than the threshold, trigger the retraining process of the model, L(t, f) represents the time-frequency feature function of the load, H is used to adjust the filter parameters according to the time-frequency characteristics of the load, G is used to compensate for the influence of environmental factors on the load characteristics, μ, σ 2 respectively represent the mean and variance of the load characteristics. <于

[0109] When the value of L(t,f) is large, it indicates that the time-frequency characteristics of the load are consistent with the model prediction, and the identification accuracy is high. When the value of L(t,f) is small, it indicates that the time-frequency characteristics of the load differ significantly from the model prediction, and the identification accuracy is low, requiring the triggering of an adaptive update process. When the time-frequency characteristics of the load are abnormal, time-frequency feature analysis is performed, and the filter parameters in the feature extraction module are automatically adjusted to enhance the extraction of frequency band features. When the load characteristics change significantly over time, the adaptive update of the classification learning model is automatically triggered to maintain the identification accuracy. The optimized model is then reapplied to the load identification task to continue monitoring changes in load characteristics and environmental factors. When the adaptive update conditions are met, the next round of optimization process begins.

[0110] The adaptive update conditions include setting a retraining process for the model when the accuracy of load feature extraction is lower than 'a', so as to ensure that the model can adapt to the new load characteristics, and updating the model every 24 hours to improve the model's recognition.

[0111] 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.

[0112] Example 2

[0113] Reference Figures 1-2 As an embodiment of the present invention, a nonlinear load identification method based on time-frequency analysis is provided. To verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0114] The test used total power data from 56 households selected from the publicly available PLAID dataset containing high-frequency data. The training and testing sets for the model were randomly allocated in a 3:1 ratio. Evaluation metrics employed included state identification accuracy metrics commonly used in the field of non-intrusive load monitoring, such as accuracy, precision, recall, and F1 score.

[0115] The key parameters set in the embodiment are as follows: the initial sliding window length N is 100, the number of hidden layer units L in the extreme learning machine model is 10, and the target accuracy τ is 93%.

[0116] Training and testing were conducted according to the above experimental configuration, and the test results are shown in Table 1.

[0117] Table 1 Test Results

[0118]

[0119]

[0120] The results show that the identification method described in this invention significantly outperforms the non-invasive load monitoring field in identifying various electrical appliances, and performs exceptionally well in some common electrical appliances, such as hair dryers, heaters, incandescent lamps, and vacuum cleaners.

[0121] 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.

[0122] Example 3

[0123] The third embodiment of the present invention differs from the first two embodiments in that:

[0124] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0126] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0127] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0128] Example 4

[0129] Reference Figure 3 This is the fourth embodiment of the present invention. This embodiment provides a nonlinear load identification system based on time-frequency analysis, including a data acquisition module, a multi-transient load feature extraction module, an extreme learning machine network construction module, a model training and adjustment module, and a performance evaluation module.

[0130] The data acquisition module collects current time series data from the user's main outlet to obtain the raw current signal of the nonlinear load, providing data support for subsequent feature extraction and model training.

[0131] The multi-transient load feature extraction module extracts time-frequency features from the collected current time series data, including frequency domain amplitude, frequency domain phase, instantaneous amplitude, instantaneous phase, and instantaneous frequency, transforming the original current signal into easily identifiable feature vectors and improving the accuracy of load identification.

[0132] The Extreme Learning Machine Network Construction Module constructs an Extreme Learning Machine network containing an input layer, hidden layers, and an output layer. It effectively fuses the multi-time-frequency domain features output by the feature extraction module and begins subsequent model training.

[0133] The model training and tuning module trains the Extreme Learning Machine network, adjusts the network weights and biases to achieve good performance on the training data, performs 5-Fold cross-validation on the model, uses each subset as a test set to evaluate the model, and adjusts the model based on the results.

[0134] The performance evaluation module evaluates the performance of the trained model on the test set, quantifies the performance of the entire load identification system, and provides a basis for model optimization and adjustment.

[0135] 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 nonlinear load identification method based on time-frequency analysis, characterized in that: include, Data is collected and preprocessed. The target time series is extracted by a sliding window through a multi-transient load feature extraction module. Time-frequency domain features are extracted based on fast Fourier transform and Hilbert-Huang transform, respectively. Based on the extracted time-domain and frequency-domain multi-transient load features, a multi-feature fusion classification learning model is established. The nonlinear load identification was trained and adjusted using a public dataset. The sliding window length was continuously adjusted based on the extracted time-frequency feature performance and the load identification performance to obtain the optimal test results and to optimize the results in combination with load characteristics.

2. The nonlinear load identification method based on time-frequency analysis as described in claim 1, characterized in that: The data acquisition and preprocessing process includes collecting data through current sensors, voltage sensors, temperature sensors, and data acquisition devices; preprocessing the raw data including denoising, normalization, and synchronization; extracting features from different modal data and encoding them into high-dimensional vectors; and comprehensively analyzing load characteristics under different modalities through a multimodal fusion strategy based on an attention mechanism: the feature vectors of different modalities are input into an attention layer, which calculates the correlation between different features using a weight matrix to obtain a weight vector; the weight vector is then weighted and summed with the feature vectors of each modality to obtain a fused feature vector; and the fused feature vector is input into a learning model for nonlinear load identification.

3. The nonlinear load identification method based on time-frequency analysis as described in claim 2, characterized in that: The extraction of time-frequency domain features includes truncating the target time series by the sliding window length and extracting features from the truncated sequence. Feature extraction includes time-domain feature extraction and frequency-domain feature extraction to obtain complete feature samples to the greatest extent possible. The initial sliding window length is defined as N. Based on this window length, the total current signal of the user is truncated, resulting in the target current time series as follows: {i(n)|n=1,2,…,N-1,N} The frequency domain feature extraction includes extracting frequency domain features, including frequency domain amplitude FA and frequency domain phase FP, from the target current time series using short-time Fourier transform. The original current signal is processed by short-time Fourier transform as follows: The frequency domain amplitude FA and frequency domain phase FP are calculated as follows: FR(m)=arctan(ImF(m) / ReF(m)) Where i(n) is the instantaneous current value at each time point n of the current time series sample, F(m) is the frequency domain signal of the original current signal i(n) of the user's total port, Re(.) and Im(.) are the real part and imaginary part of the target, respectively, m represents the harmonic order, and j is the imaginary unit.

4. The nonlinear load identification method based on time-frequency analysis as described in claim 3, characterized in that: The time-domain feature extraction includes processing the acquired user port raw current signal i(n) using Hilbert-Huang transform, and the time-domain features include instantaneous amplitude IA, instantaneous phase IP, and instantaneous frequency IF; The Hilbert-Huang transform processing of the original current signal i(n) at the main port is as follows: After Hilbert-Huang transform processing, the original current signal i(t) is analyzed and characterized, and its signal form is analyzed as follows: The two time-domain features extracted, instantaneous amplitude IA and instantaneous phase IP, are shown below: IP(t) = arctan(HT(i(t)) / i(t)) After obtaining the instantaneous amplitude IA and instantaneous phase IP, the following calculations are performed: Where j represents the imaginary unit, This indicates that the first derivative is obtained in time. Since the obtained characteristic instantaneous phase IP is a discrete time series in form, the derivative result is the difference between two data points, and u represents the integrand variable.

5. The nonlinear load identification method based on time-frequency analysis as described in claim 4, characterized in that: The classification learning model includes an extreme learning machine model as the main body, which completes the identification of the load by fusing multiple features of the learning input. The extreme learning machine consists of an output layer, an input layer and multiple hidden layers. In the structure of the network, the weight values ​​and bias values ​​from the input layer to the hidden layer are randomly generated and are not iteratively modified. Output value t of output layer node k With input layer node value x i The mathematical relationship is as follows: When ω is determined j and b j Next, calculate the output matrix H of the hidden layer: The time-frequency domain features are fused to complete the subsequent model training. A fully connected network is adopted, and the results are used to train the established extreme learning machine model. Where 1 represents the hidden layer number in the network, ω l and b l These are the weight matrix and bias vector from the input layer to the first hidden layer node in the network structure, respectively. g(·) refers to the activation function, and β... l This refers to the weight values ​​from the first hidden layer node to the output layer node, L refers to the number of hidden layer nodes, and T refers to the output matrix of the output layer. Let H be the pseudo-inverse matrix.

6. The nonlinear load identification method based on time-frequency analysis as described in claim 5, characterized in that: The training and adjustment of nonlinear load identification includes continuously adjusting the sliding window length based on the extracted time-frequency features and load identification performance at each window length, using the 5-Fold algorithm for cross-validation, and adjusting the weighting parameters R of the feature fusion connection network. When E avg (W new )≥E and E avg (W new )-E avg (W old When )≥ΔE, accept the new weighting parameter, when E avg (W new ) < E or E avg (W new )-E avg (W old If ) < ΔE, reject the new weighting parameter; The original subset is randomly and evenly divided into five subsets. One subset is selected as the test set, and the other four subsets are used as the training set. After training the model, it is evaluated on the test set. Accuracy and precision are used as performance evaluation metrics. The training and testing of the model are repeated five times, with each subset used as the test set in turn for verification. This ensures that each subset is used as a test set for model evaluation. The average of the performance metrics is taken as the final result. The weight parameters are adjusted to make the performance metrics after cross-validation the best and to meet the corresponding target accuracy τ. After cross-validation, it is verified whether the target accuracy value has been reached. If the target accuracy value has been reached, the final result is output and the training ends. If the target accuracy value has not been reached, the parameters are adjusted and feature fusion and training are repeated until the target accuracy value is reached and the training ends. Among them, W opt X is a weighted parameter optimized by minimizing the average loss and the exponential decay of accuracy. i Let L(F(x; W), y) represent the i-th subset of X, and L(F(x; W), y) measure the difference between the predicted F(x; W) and the true label y. new It is the adjusted weighted parameter, W. o1d It is the original weighting parameter, E avg For average performance measurement, ΔE is the performance improvement threshold, and E is the set performance threshold.

7. The nonlinear load identification method based on time-frequency analysis as described in claim 6, characterized in that: The optimization based on load characteristics includes the model automatically adjusting its internal structure and parameters according to the time-frequency characteristics of the load to ensure that the identification effect is always in the optimal state, and monitoring and compensating for the impact of environmental factors on load characteristics. Where 'a' represents the threshold for load feature extraction accuracy; when it falls below the threshold, the model retraining process is triggered. L(t, f) represents the time-frequency characteristic function of the load. H is used to adjust the filter parameters based on the load's time-frequency characteristics. G is used to compensate for the influence of environmental factors on load characteristics. μ and σ... 2 These represent the mean and variance of the load characteristics, respectively. When the value of L(t,f) is large, it indicates that the time-frequency characteristics of the load are consistent with the model prediction and the identification accuracy is high. When the value of L(t,f) is small, it indicates that there is a large difference between the time-frequency characteristics of the load and the model prediction, the identification accuracy is low, and an adaptive update process needs to be triggered. When the time-frequency characteristics of the load are abnormal, time-frequency characteristic analysis is performed, and the filter parameters in the feature extraction module are automatically adjusted to enhance the extraction of frequency band features. When the load characteristics change significantly over time, the adaptive update of the classification learning model is automatically triggered to maintain the identification accuracy. The optimized model is reapplied to the load identification task to continue monitoring changes in load characteristics and environmental factors. When the adaptive update conditions are met, the next round of optimization process begins. The adaptive update conditions include setting a retraining process for the model when the accuracy of load feature extraction is lower than 'a', so as to ensure that the model can adapt to the new load characteristics, and updating the model every 24 hours to improve the model's recognition.

8. A system employing a nonlinear load identification method based on time-frequency analysis as described in any one of claims 1 to 7, characterized in that: It includes a data acquisition module, a multi-transient load feature extraction module, an extreme learning machine network construction module, a model training and tuning module, and a performance evaluation module; The data acquisition module collects current time series data from the user's main outlet to obtain the raw current signal of the nonlinear load, providing data support for subsequent feature extraction and model training. The multi-transient load feature extraction module performs time-frequency feature extraction on the collected current time series data, including frequency domain amplitude, frequency domain phase, instantaneous amplitude, instantaneous phase, and instantaneous frequency, transforming the original current signal into an easily identifiable feature vector, thereby improving the accuracy of load identification. The Extreme Learning Machine Network Construction Module constructs an Extreme Learning Machine Network containing an input layer, a hidden layer, and an output layer, effectively fuses the multi-time-frequency domain features output by the feature extraction module, and begins subsequent model training. The model training and tuning module trains the Extreme Learning Machine network, adjusts the network weights and biases to achieve good performance on the training data, performs 5-Fold cross-validation on the model, uses each subset as a test set to evaluate the model, and adjusts it based on the results. The performance evaluation module evaluates the performance of the trained model on the test set, quantifies the performance of the entire load identification system, and provides a basis for model optimization and adjustment.

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 method according to any one of claims 1 to 5.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images