Non-intrusive load monitoring method and system based on transient-state images

Abstract

The application provides a kind of non-invasive load monitoring method and system based on transient state image, the method comprises: judging whether there is load switching event;If there is load switching event, with the time point of event occurrence as the first reference point, to screen out pre-event steady state window and post-event steady state window, and extract steady state feature and electrical feature;Transient window is constructed, and transient feature is extracted according to transient window;Steady state feature and transient feature are respectively imaged, and steady state image, transient image, electrical feature are input into three flow parallel deep learning network, the probability under each identification result is obtained, and the identification result corresponding to the maximum value of probability is selected from the probability of all identification results as output result.The application can improve the accuracy of load monitoring.

Description

Technical Field

[0001] This invention relates to the field of load monitoring technology, and in particular to a non-intrusive load monitoring method and system based on transient steady-state images. Background Technology

[0002] With the rapid development of smart grids and the energy internet, IoT meters, as core sensing devices on the end-user side, have evolved from traditional electricity metering tools into intelligent terminals with bidirectional communication, real-time data acquisition, and edge computing capabilities. Through IoT technology, they enable remote transmission and centralized management of electricity consumption data, providing a data foundation for user energy efficiency analysis, demand response, and grid optimization scheduling. However, load monitoring requires intrusive sub-meter installations to obtain individual load information, which presents problems such as high deployment costs, complex maintenance, and significant user privacy risks. Therefore, non-intrusive load monitoring technology has emerged. This technology only needs to analyze the total electricity consumption signal collected by the IoT meter to decouple and identify the operating status of individual internal appliances, becoming an important direction in smart electricity research.

[0003] Non-intrusive load monitoring (NILM) is a technology that analyzes the total current and voltage waveforms at the power inlet to decompose and identify the types and operating states of various internal electrical loads. Due to its ease of installation and low cost, this technology has broad application prospects in energy management and power diagnostics. Traditional NILM research methods are mostly based on data with low sampling rates, and the features utilized are mainly concentrated on steady-state power, VI trajectory, and harmonics. These methods have a certain identification effect on loads with large differences in characteristics; however, when faced with the diverse types and complex characteristics of modern household appliances, especially when the steady-state characteristics of many loads tend to be similar, the identification accuracy and reliability of traditional methods face significant challenges.

[0004] With the development of high-frequency data acquisition technology, rich transient details during load switching have been captured, providing a new data foundation for improving identification accuracy. However, existing research utilizing high-frequency data still has significant limitations: First, most methods focus only on single-type features, such as using only the amplitude sequence of transient current or steady-state waveforms, failing to comprehensively mine the multi-dimensional information contained in high-frequency data, including steady-state, transient, and quantization features, resulting in insufficient information utilization; second, feature fusion methods are mostly simple fixed-weighted scalars or decision-level voting, failing to dynamically adjust the confidence of different feature sources according to the characteristics of specific load switching events, resulting in insufficient adaptability and intelligence of the fusion strategy, limiting further improvement in model performance. Summary of the Invention

[0005] The purpose of this invention is to provide a non-intrusive load monitoring method and system based on transient steady-state images. This aims to solve the problem that traditional technologies focus on a single type of feature, such as only using the amplitude sequence of transient current or steady-state waveform, and fail to comprehensively mine the multi-dimensional information such as steady-state, transient and quantization features contained in high-frequency data, resulting in insufficient information utilization and affecting the accuracy of load monitoring.

[0006] In a first aspect, the present invention provides a non-intrusive load monitoring method based on transient steady-state images, the method comprising: Obtain the current RMS value sequence under continuous power frequency cycle, and determine whether there is a load switching event based on the current RMS value sequence; If a load switching event exists, the time point of the event occurrence is used as the first reference point. Based on the first reference point, a steady-state window before the event and a steady-state window after the event are selected. Steady-state features and electrical features are extracted based on the first reference point, the steady-state window before the event, and the steady-state window after the event. The transient window is constructed by taking the starting point of the steady-state window before the event as the starting point of the transient window and the ending point of the steady-state window after the event as the ending point of the transient window, and transient features are extracted based on the transient window. The steady-state features and the transient features are respectively visualized to obtain steady-state images and transient images. The steady-state images, the transient images, and the electrical features are then input into a three-stream parallel deep learning network to obtain the probability of each identification result. Finally, the identification result corresponding to the maximum probability is selected from all the probabilities of the identification results as the output result.

[0007] In some embodiments, the step of obtaining the current RMS value sequence under continuous power frequency cycles and determining whether a load switching event exists based on the current RMS value sequence includes: The current intensity for each power frequency cycle can be obtained using the following formula: ; in, To correspond to the current intensity under the power frequency cycle, The number of sampling points per power frequency cycle. Represents the first in a period Current values ​​at each sampling point; If the following formula is satisfied, then a load switching event is determined to exist: ; in, , The first j +1, the first j Current intensity per cycle This is the difference in current intensity. This is a dynamic threshold.

[0008] In some embodiments, if a load switching event exists, the step of using the event occurrence time as a first reference point, filtering out a pre-event steady-state window and a post-event steady-state window based on the first reference point, and extracting steady-state features and electrical features based on the first reference point, the pre-event steady-state window, and the post-event steady-state window includes: Search forward M power frequency cycles along the first reference point with a preset window. If the power intensity value no longer rises or falls continuously, the preset window at this time is determined to be the pre-event steady state window, and the first sampling point of the second in the positive order of all complete power frequency cycles contained in the pre-event steady state window is taken as the second reference point. Search M power frequency cycles backward along the first reference point using the preset window. If the power intensity value no longer rises or falls continuously, then the preset window at this time is determined to be the post-event steady-state window, and the second-to-last first sampling point in all complete power frequency cycles contained in the post-event steady-state window is taken as the third reference point. The first zero-crossing point is the time point closest to the second reference point in terms of time length and when the voltage signal changes from negative to positive. The second zero-crossing point is the time point closest to the third reference point in terms of time length and when the voltage signal changes from negative to positive. The current waveform of a complete power frequency cycle after the first zero-crossing point is extracted as the current waveform of the steady state before the event, and the current waveform of a complete power frequency cycle after the second zero-crossing point is extracted as the current waveform of the steady state after the event. The steady-state current waveform during load switching events can be obtained using the following formula: ; in, The steady-state current waveform during load switching events. This is the steady-state current waveform after the event. The current waveform in the steady state before the event; The preset window contains multiple power frequency cycles, and the time period it contains is less than the time period of M cycles.

[0009] In some embodiments, if a load switching event exists, the step of using the event occurrence time as a first reference point to filter out a pre-event steady-state window and a post-event steady-state window based on the first reference point, and extracting steady-state features and electrical features based on the first reference point, the pre-event steady-state window, and the post-event steady-state window further includes: Extraction and The voltage waveform corresponding to a complete cycle is used as the single-cycle voltage waveform of the load switching event. Electrical characteristics are extracted based on the single-cycle voltage waveform and the steady-state current waveform of the load switching event. The electrical characteristics include: active power, peak current coefficient, reactive power, correlation coefficient, apparent power, RMS current, power factor, DC component, power factor angle, peak current area, current amplitude, duty cycle, current distortion rate, fifth current harmonic, third current harmonic, seventh current harmonic, and fundamental current.

[0010] In some embodiments, the step of constructing a transient window by taking the starting point of the pre-event steady-state window as the starting point of the transient window and the ending point of the post-event steady-state window as the ending point of the transient window, and extracting transient features based on the transient window includes: Extract the current intensity of each complete power frequency cycle within the transient window to construct the original transient current intensity sequence; Calculate the first average value of the current intensity across all complete power frequency cycles contained in the steady-state window before the event, and subtract the first average value from the original transient current intensity sequence to obtain the transient change sequence caused by the pure load switching event.

[0011] In some embodiments, the electrical characteristic further includes a fluctuation index, which is calculated according to the following formula: ; ; ; in, As an indicator of volatility, The standard deviation of the current intensity in the steady-state window after the event. The standard deviation of the current intensity within the steady-state window prior to the event. This represents the average current intensity during the steady-state window after the event. The mean current intensity during the steady-state window before the event. The standard deviation of the current intensity within the steady-state window, which includes the post-event steady-state window and the pre-event steady-state window. The mean current intensity during the steady-state window. This represents the number of complete power frequency cycles contained within the steady-state window.

[0012] In some embodiments, the step of imagerizing the steady-state features and the transient features to obtain steady-state images and transient images includes: The steady-state current waveform of the load switching event is converted into a standardized two-dimensional grayscale image to obtain the steady-state image; Create a blank RGB image with a size of 128×64 pixels and fill the background with white. Linearly map all sampling points of the transient change sequence to the entire width of the blank RGB image, and linearly map the amplitude range of the transient change sequence to the entire height of the blank RGB image. The difference between the maximum and minimum values ​​in the transient change sequence, the number of complete power frequency cycles contained in the transient window, and the fluctuation index are normalized to the interval [0, 1] to map to the range [0, 255] of the RGB channel, thus obtaining the transient image.

[0013] In some embodiments, the step of inputting the steady-state image, the transient image, and the electrical features into a three-stream parallel deep learning network to obtain the probability of each identification result, and selecting the identification result corresponding to the maximum probability from all the probabilities of the identification results as the output result includes: The steady-state image stream is processed using a lightweight CNN structure that performs parallel extraction and fusion of multi-scale features. The transient sequence image stream employs a deep fusion network that introduces a cross-layer feature reuse mechanism to process transient images; The feature stream uses a multilayer perceptron based on a gated attention mechanism to process 18-dimensional electrical features; The probability of each identification result is obtained using the following formula: ; in, This represents a probability distribution, including the probability of each identification outcome. , , All are weights. , , These are the outputs of the steady-state image stream, the transient sequence image stream, and the feature stream, respectively.

[0014] Secondly, the present invention provides a non-intrusive load monitoring system based on transient steady-state images, the system comprising: The switching event detection module is used to acquire the current effective value sequence under continuous power frequency cycle, and determine whether there is a load switching event based on the current effective value sequence; The first feature extraction module is used to, if there is a load switching event, take the time of the event as the first reference point, filter out the steady-state window before the event and the steady-state window after the event based on the first reference point, and extract steady-state features and electrical features based on the first reference point, the steady-state window before the event and the steady-state window after the event. The second feature extraction module is used to construct a transient window with the starting point of the steady-state window before the event as the starting point of the transient window and the ending point of the steady-state window after the event as the ending point of the transient window, and to extract transient features based on the transient window. The probability prediction module is used to visualize the steady-state features and the transient features respectively to obtain steady-state images and transient images, and input the steady-state images, the transient images, and the electrical features into a three-stream parallel deep learning network to obtain the probability of each identification result, and select the identification result corresponding to the maximum probability from all the probabilities of the identification results as the output result.

[0015] Thirdly, the present invention provides a storage medium that stores one or more programs that, when executed by a processor, implement the above-described non-intrusive load monitoring method based on a transient steady-state image.

[0016] Fourthly, the present invention provides an electronic device, the electronic device comprising a memory and a processor, wherein: The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the above-described non-intrusive load monitoring method based on transient steady-state images.

[0017] Compared with the prior art, the present invention has the following advantages: This invention accurately determines load switching events by acquiring the effective current value sequence under continuous power frequency cycles; it filters steady-state windows and extracts steady-state and electrical features based on the event occurrence time, while simultaneously constructing transient windows to extract transient features, comprehensively capturing load change information; it visualizes steady-state and transient features, combines them with electrical features, and inputs them into a three-stream parallel deep learning network. Utilizing multi-scale feature extraction, cross-layer feature reuse, and gated attention mechanisms, it efficiently fuses different feature information, accurately calculates the probability of each identification result, and outputs the identification result corresponding to the highest probability, effectively improving the accuracy and reliability of load monitoring and providing strong support for the refined management of power systems. Attached Figure Description

[0018] Figure 1 This is a flowchart of a non-intrusive load monitoring method based on transient steady-state images proposed in an embodiment of the present invention; Figure 2 This is a schematic diagram of steady-state and transient windows; Figure 3 (a) and (b) in the figure are schematic diagrams of steady-state and transient images, respectively; Figure 4 This is a schematic diagram of a non-intrusive load monitoring system based on transient steady-state images proposed in an embodiment of the present invention.

[0019] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.

[0021] like Figure 1 As shown, an embodiment of the present invention proposes a non-intrusive load monitoring method based on transient steady-state images. The method includes steps S101 to S104, wherein: Step S101: Obtain the current effective value sequence under continuous power frequency cycle, and determine whether there is a load switching event based on the current effective value sequence; It should be noted that in this step, a non-intrusive monitoring terminal is first deployed at the user's power inlet to collect total current and voltage signals at a frequency of 10kHz. Each 50Hz power frequency cycle contains 100 data points. Then, the current intensity in each power frequency cycle is obtained according to the following formula: ; in, To correspond to the current intensity under the power frequency cycle, The number of sampling points per power frequency cycle. Represents the first in a period Current values ​​at each sampling point; If the following formula is satisfied, then a load switching event is determined to exist; otherwise, a load switching event does not exist: ; in, , The first j +1, the first j Current intensity per cycle This is the difference in current intensity. This is a dynamic threshold.

[0022] Step S102: If a load switching event exists, the time point of the event occurrence is taken as the first reference point. The steady-state window before the event and the steady-state window after the event are selected based on the first reference point. Steady-state features and electrical features are extracted based on the first reference point, the steady-state window before the event, and the steady-state window after the event. It should be noted that steady-state characteristics describe the current change characteristics when a single load is running. Some loads, due to their unique electrical structure, exhibit a consistent waveform shape that is clearly distinct from a sine wave. Therefore, single-cycle current waveform images can also provide effective identification criteria.

[0023] Based on this, in some embodiments, firstly, a preset window is used to search forward M power frequency cycles along the first reference point. If the power intensity value no longer rises or falls continuously, then the preset window at this time is determined to be a pre-event steady-state window, and the first sampling point of the second-to-last complete power frequency cycle contained in the pre-event steady-state window is used as the second reference point. Then, the preset window is used to search backward M power frequency cycles along the first reference point. If the power intensity value no longer rises or falls continuously, then the preset window at this time is determined to be a post-event steady-state window, and the first sampling point of the second-to-last complete power frequency cycle contained in the post-event steady-state window is used as the third reference point.

[0024] Then, the time point closest to the second reference point in time length and where the voltage signal changes from negative to positive is taken as the first zero-crossing point, and the time point closest to the third reference point in time length and where the voltage signal changes from negative to positive is taken as the second zero-crossing point. The current waveform of a complete power frequency cycle after the first zero-crossing point is extracted as the current waveform of the steady state before the event, and the current waveform of a complete power frequency cycle after the second zero-crossing point is taken as the current waveform of the steady state after the event. The steady-state current waveform during load switching events can be obtained using the following formula: ; in, The steady-state current waveform during load switching events. This is the steady-state current waveform after the event. The current waveform represents the steady-state current waveform before the event.

[0025] Furthermore, in some embodiments, the preset window includes multiple power frequency cycles, and the time period it includes is less than the time period of M cycles.

[0026] In addition, in some embodiments, it is also necessary to extract and The voltage waveform corresponding to a complete cycle is used as the single-cycle voltage waveform of the load switching event. Electrical features are extracted based on the single-cycle voltage waveform and the steady-state current waveform of the load switching event. These electrical features include: active power, peak current coefficient, reactive power, correlation coefficient, apparent power, RMS current, power factor, DC component, power factor angle, peak current area, current amplitude, duty cycle, current distortion rate, fifth harmonic, third harmonic, seventh harmonic, and fundamental current. The correlation coefficient is an index used to quantify the linear similarity in shape between the normalized single-cycle current waveform and an ideal sine wave. When the coefficient is 1, it indicates that the current waveform is completely consistent with the ideal sine wave; the closer the coefficient is to 0, the more severe the waveform distortion and the lower the similarity. This correlation coefficient is obtained by normalizing a one-cycle current waveform and a standard sine wave of the same frequency, and then calculating their linear correlation. The specific calculations of the remaining features are conventional techniques and will not be described in detail in this embodiment.

[0027] In addition, in some embodiments, the electrical characteristics also include a fluctuation index. Since some loads (such as frequency converters and motor loads) have inherent fluctuations in current even in steady state, this is an important identification feature. In order to quantify the natural fluctuation of the effective value of current when the load is running in steady state, the fluctuation index is defined as the fluctuation index.

[0028] The current RMS value sequence before and after the event is extracted. The volatility index is the ratio of the standard deviation to the mean, a dimensionless relative volatility measure that eliminates the influence of different load current amplitudes, making the volatility of different power loads comparable. The difference between the relative volatility measures before and after the event is the volatility index. ; ; ; in, As an indicator of volatility, The standard deviation of the current intensity in the steady-state window after the event. The standard deviation of the current intensity within the steady-state window prior to the event. This represents the average current intensity during the steady-state window after the event. The mean current intensity during the steady-state window before the event. The standard deviation of the current intensity within the steady-state window, which includes the post-event steady-state window and the pre-event steady-state window. The mean current intensity during the steady-state window. This represents the number of complete power frequency cycles contained within the steady-state window.

[0029] Step S103: Construct a transient window by taking the starting point of the steady-state window before the event as the starting point of the transient window and the ending point of the steady-state window after the event as the ending point of the transient window, and extract transient features based on the transient window; In this step, such as Figure 2 As shown, the starting point of the steady-state window before the event is the starting point of the transient window, and the ending point of the steady-state window after the event is the ending point of the transient window. The transient time window is all the cycles between the starting point and the ending point of the transient window. This window completely covers the entire dynamic process from the old steady state to the new steady state. Then, the current intensity of each complete power frequency cycle within the transient window is extracted to form the original transient current intensity sequence; then, the first average value of the current intensity of all complete power frequency cycles included in the steady-state window before the event is calculated, and the original transient current intensity sequence is subtracted from the first average value to obtain the transient change sequence caused by the pure load switching event. The core purpose of this step is to eliminate the influence of the base load and make the transient change sequence reflect the transient characteristics of the switched load itself, with its initial value close to zero.

[0030] Step S104: The steady-state features and the transient features are visualized to obtain steady-state images and transient images, and the steady-state images, the transient images, and the electrical features are input into a three-stream parallel deep learning network to obtain the probability of each identification result, and the identification result corresponding to the maximum probability is selected from the probabilities of all identification results as the output result.

[0031] It should be noted that, as Figure 3 As shown in (a) and (b), the steady-state current waveform of the load switching event is converted into a standardized two-dimensional grayscale image to obtain the steady-state image. Specifically, the first step is to... Maximum absolute value normalization is performed to eliminate the influence of load power magnitude on image amplitude, allowing the model to focus on learning waveform shape features; then comes image generation and size normalization. The normalized single-cycle current waveform is mapped to a 64x64 pixel image with a white background. Each point is mapped to coordinates on the canvas and drawn using black. To obtain a clearer image, linear interpolation is performed between adjacent points to draw line segments instead of just drawing discrete points. The final result is a standard-sized grayscale image that clearly shows the unique current waveform shape after load normalization, which can be used as input for subsequent image recognition networks.

[0032] Furthermore, in some embodiments, the process of acquiring the transient image is as follows: a blank RGB image of size 128×64 pixels is created, with the background filled with white. All sampling points of the transient change sequence are linearly mapped to the entire width direction of the blank RGB image, and the amplitude range of the transient change sequence is linearly mapped to the entire height direction of the blank RGB image. Then, the difference between the maximum and minimum values ​​in the transient change sequence, the number of complete power frequency cycles contained in the transient window, and the fluctuation index are normalized to the [0, 1] interval to map to the [0, 255] range of the RGB channels. For each point in the transient change sequence, a unique color is calculated for each point. In order to form a continuous and smoothly changing color trajectory, linear interpolation is performed between adjacent points, and the corresponding color is calculated proportionally for each interpolation point, and then drawn. Finally, a standard-sized color image is obtained. This visualization method enables the convolutional neural network to indirectly process these key physical features while extracting image spatial features, achieving effective fusion of multimodal information and enhancing the model's recognition ability and interpretability.

[0033] Furthermore, in some embodiments, a three-stream parallel deep learning network is constructed to process steady-state images, transient images, and one-dimensional electrical features respectively. The outputs of the three streams are integrated through an adaptive fusion module to ultimately output a high-precision load identification result. This design fully considers the differences in the dependence of different load categories on different feature streams, achieving intelligent fusion decision-making.

[0034] Specifically, the steady-state image stream employs a lightweight CNN structure that combines multi-scale feature extraction and fusion to process steady-state images. At the beginning of the network, multiple convolutional kernels of different sizes are used in parallel, and the resulting feature maps are then concatenated. This allows the network to simultaneously capture subtle local distortions and the overall shape contour of the current waveform. Finally, through fully connected layers and softmax layers, the network outputs a load category probability distribution based on the steady-state waveform image. .

[0035] The transient image stream employs a deep fusion network that introduces a cross-layer feature reuse mechanism to process transient image sequences. The main body of the network adopts the densely connected structure of DenseNet. The input to each convolutional layer comes not only from the previous layer but also from the feature maps of all preceding layers. This facilitates the propagation and reuse of features within the network, which is particularly effective for learning complex color-shape joint features during long transient processes. At the end of the network, before the global average pooling layer, a bidirectional long short-term memory network module is added. This module treats the spatial feature sequence extracted by the convolutional network as a time series, thereby explicitly modeling the temporal dependencies of the transient process. The network finally outputs a probability distribution of the load category based on the transient color image. .

[0036] The one-dimensional electrical feature stream uses a multilayer perceptron based on a gated attention mechanism to process 18-dimensional electrical features, including 17 common electrical features and one fluctuation index. The input features are first passed through a gating unit, a single-layer neural network, which outputs an 18-dimensional gating vector G with values ​​between [0,1]. Soft selection of the input features is then performed through element-wise multiplication. ; in This is the Sigmoid function. This operation can suppress noise or irrelevant features and enhance important features.

[0037] Gated features Input a self-attention layer. This layer computes the relationships between 18 features and generates an attention weight for each feature. The final weighted eigenvector is ; This step enables the model to dynamically focus on the most discriminative feature combinations in the current sample, The network takes a standard MLP as input for classification and outputs a load category probability distribution based on one-dimensional electrical features. .

[0038] The probability distribution of the three stream outputs uses a dynamic weight generation method instead of fixed weights. It not only obtains the output probabilities of the three streams but also extracts the feature vectors from the penultimate layer (i.e., the Softmax layer). These feature vectors contain richer information about the model's decision confidence than the final probabilities.

[0039] The three feature vectors are concatenated into a single feature vector, Vconcat. Vconcat is then input into a neural network consisting of two fully connected layers, which ultimately outputs a 3-dimensional weight vector. ; This mapping function can dynamically allocate the most suitable fusion weights based on the feature confidence level exhibited by the current sample in the three-stream model. For example, for a load with a very unique steady-state waveform, the meta-network will... Assign higher weights; for a load that is extremely significant for a transient process, The weight will be higher.

[0040] The final probability distribution is obtained by weighting the three stream outputs according to their adaptive weights: ; in, This represents a probability distribution, including the probability of each identification outcome. , , All are weights. , , These are the outputs of the steady-state image stream, the transient sequence image stream, and the feature stream, respectively.

[0041] In summary, this invention accurately determines load switching events by acquiring the effective current value sequence under continuous power frequency cycles; it filters steady-state windows and extracts steady-state and electrical features based on the event occurrence time, while simultaneously constructing transient windows to extract transient features, comprehensively capturing load change information; it visualizes steady-state and transient features, combines them with electrical features, and inputs them into a three-stream parallel deep learning network. Utilizing multi-scale feature extraction, cross-layer feature reuse, and gated attention mechanisms, it efficiently fuses different feature information, accurately calculates the probability of each identification result, and outputs the identification result corresponding to the highest probability, effectively improving the accuracy and reliability of load monitoring and providing strong support for the refined management of power systems.

[0042] like Figure 4 As shown, one embodiment of the present invention proposes a non-intrusive load monitoring system based on transient steady-state images, the system comprising: The switching event detection module 10 is used to acquire the current effective value sequence under continuous power frequency cycle, and determine whether there is a load switching event based on the current effective value sequence. The first feature extraction module 20 is used to, if there is a load switching event, take the time point of the event as the first reference point, filter out the steady-state window before the event and the steady-state window after the event based on the first reference point, and extract steady-state features and electrical features based on the first reference point, the steady-state window before the event and the steady-state window after the event. The second feature extraction module 30 is used to construct a transient window with the starting point of the steady-state window before the event as the starting point of the transient window and the ending point of the steady-state window after the event as the ending point of the transient window, and to extract transient features based on the transient window. The probability prediction module 40 is used to visualize the steady-state features and the transient features respectively to obtain steady-state images and transient images, and input the steady-state images, the transient images, and the electrical features into a three-stream parallel deep learning network to obtain the probability of each identification result, and select the identification result corresponding to the maximum probability from all the probabilities of the identification results as the output result.

[0043] In another aspect, the present invention also proposes a storage medium on which one or more programs are stored, which, when executed by a processor, implement the above-described non-intrusive load monitoring method based on metastatic images.

[0044] In another aspect, the present invention also proposes an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to realize the above-mentioned non-intrusive load monitoring method based on transient steady-state images.

[0045] Those skilled in the art will understand that 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 mean any means that can contain stored, communicated, propagated, or transmitted programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0046] 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.

[0047] 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.

[0048] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. A non-intrusive load monitoring method based on transient steady-state images, characterized in that, The method includes: Obtain the current RMS value sequence under continuous power frequency cycle, and determine whether there is a load switching event based on the current RMS value sequence; If a load switching event exists, the time point of the event occurrence is used as the first reference point. Based on the first reference point, a steady-state window before the event and a steady-state window after the event are selected. Steady-state features and electrical features are extracted based on the first reference point, the steady-state window before the event, and the steady-state window after the event. The transient window is constructed by taking the starting point of the steady-state window before the event as the starting point of the transient window and the ending point of the steady-state window after the event as the ending point of the transient window, and transient features are extracted based on the transient window. The steady-state features and the transient features are respectively visualized to obtain steady-state images and transient images. The steady-state images, the transient images, and the electrical features are then input into a three-stream parallel deep learning network to obtain the probability of each identification result. Finally, the identification result corresponding to the maximum probability is selected from all the probabilities of the identification results as the output result.

2. The non-intrusive load monitoring method based on transient steady-state images according to claim 1, characterized in that, The step of obtaining the current RMS value sequence under continuous power frequency cycle and determining whether a load switching event exists based on the current RMS value sequence includes: The current intensity for each power frequency cycle can be obtained using the following formula: ； in, To correspond to the current intensity under the power frequency cycle, The number of sampling points per power frequency cycle. Represents the first in a period Current values ​​at each sampling point; If the following formula is satisfied, then a load switching event is determined to exist: ； in, , The first j +1, the first j Current intensity per cycle This is the difference in current intensity. This is a dynamic threshold.

3. The non-intrusive load monitoring method based on transient steady-state images according to claim 2, characterized in that, If a load switching event exists, the step of using the event occurrence time as the first reference point, filtering out the pre-event steady-state window and the post-event steady-state window based on the first reference point, and extracting steady-state features and electrical features based on the first reference point, the pre-event steady-state window, and the post-event steady-state window includes: Search forward M power frequency cycles along the first reference point with a preset window. If the power intensity value no longer rises or falls continuously, the preset window at this time is determined to be the pre-event steady state window, and the first sampling point of the second in the positive order of all complete power frequency cycles contained in the pre-event steady state window is taken as the second reference point. Search M power frequency cycles backward along the first reference point using the preset window. If the power intensity value no longer rises or falls continuously, then the preset window at this time is determined to be the post-event steady-state window, and the second-to-last first sampling point in all complete power frequency cycles contained in the post-event steady-state window is taken as the third reference point. The first zero-crossing point is the time point closest to the second reference point in terms of time length and when the voltage signal changes from negative to positive. The second zero-crossing point is the time point closest to the third reference point in terms of time length and when the voltage signal changes from negative to positive. The current waveform of a complete power frequency cycle after the first zero-crossing point is extracted as the current waveform of the steady state before the event, and the current waveform of a complete power frequency cycle after the second zero-crossing point is extracted as the current waveform of the steady state after the event. The steady-state current waveform during load switching events can be obtained using the following formula: ； in, This is the steady-state current waveform during load switching events. This is the steady-state current waveform after the event. The current waveform in the steady state before the event; The preset window contains multiple power frequency cycles, and the time period it contains is less than the time period of M cycles.

4. The non-intrusive load monitoring method based on transient steady-state images according to claim 3, characterized in that, If a load switching event exists, the step of using the event occurrence time as the first reference point, filtering out the pre-event steady-state window and the post-event steady-state window based on the first reference point, and extracting steady-state features and electrical features based on the first reference point, the pre-event steady-state window, and the post-event steady-state window further includes: Extraction and The voltage waveform corresponding to a complete cycle is used as the single-cycle voltage waveform of the load switching event. Electrical characteristics are extracted based on the single-cycle voltage waveform and the steady-state current waveform of the load switching event. The electrical characteristics include: active power, peak current coefficient, reactive power, correlation coefficient, apparent power, RMS current, power factor, DC component, power factor angle, peak current area, current amplitude, duty cycle, current distortion rate, fifth current harmonic, third current harmonic, seventh current harmonic, and fundamental current.

5. The non-intrusive load monitoring method based on transient steady-state images according to claim 4, characterized in that, The steps of constructing a transient window by taking the starting point of the pre-event steady-state window as the starting point of the transient window and the ending point of the post-event steady-state window as the ending point of the transient window, and extracting transient features based on the transient window include: Extract the current intensity of each complete power frequency cycle within the transient window to construct the original transient current intensity sequence; Calculate the first average value of the current intensity across all complete power frequency cycles contained in the steady-state window before the event, and subtract the first average value from the original transient current intensity sequence to obtain the transient change sequence caused by the pure load switching event.

6. The non-intrusive load monitoring method based on transient steady-state images according to claim 5, characterized in that, The electrical characteristics also include a fluctuation level index, which is calculated according to the following formula: ； ； ； in, As an indicator of volatility, The standard deviation of the current intensity in the steady-state window after the event. The standard deviation of the current intensity within the steady-state window prior to the event. This represents the average current intensity during the steady-state window after the event. The mean current intensity during the steady-state window before the event. The standard deviation of the current intensity within the steady-state window, which includes the post-event steady-state window and the pre-event steady-state window. The mean current intensity during the steady-state window. This represents the number of complete power frequency cycles contained within the steady-state window.

7. The non-intrusive load monitoring method based on transient steady-state images according to claim 6, characterized in that, The step of image-encoding the steady-state features and the transient features to obtain steady-state images and transient images includes: The steady-state current waveform of the load switching event is converted into a standardized two-dimensional grayscale image to obtain the steady-state image; Create a blank RGB image with a size of 128×64 pixels and fill the background with white. Linearly map all sampling points of the transient change sequence to the entire width of the blank RGB image, and linearly map the amplitude range of the transient change sequence to the entire height of the blank RGB image. The difference between the maximum and minimum values ​​in the transient change sequence, the number of complete power frequency cycles contained in the transient window, and the fluctuation index are normalized to the interval [0, 1] to map to the range [0, 255] of the RGB channel, thus obtaining the transient image.

8. The non-intrusive load monitoring method based on transient steady-state images according to claim 7, characterized in that, The steps of inputting the steady-state image, the transient image, and the electrical features into a three-stream parallel deep learning network to obtain the probability of each identification result, and selecting the identification result corresponding to the maximum probability from all the probabilities as the output result, include: The steady-state image stream is processed using a lightweight CNN structure that performs parallel extraction and fusion of multi-scale features. The transient sequence image stream employs a deep fusion network that introduces a cross-layer feature reuse mechanism to process transient images; The feature stream uses a multilayer perceptron based on a gated attention mechanism to process 18-dimensional electrical features; The probability of each identification result is obtained using the following formula: ； in, This represents a probability distribution, including the probability of each identification outcome. , , All are weights. , , These are the outputs of the steady-state image stream, the transient sequence image stream, and the feature stream, respectively.

9. A non-intrusive load monitoring system based on transient steady-state images, characterized in that, The system includes: The switching event detection module is used to acquire the current effective value sequence under continuous power frequency cycle, and determine whether there is a load switching event based on the current effective value sequence; The first feature extraction module is used to, if there is a load switching event, take the time of the event as the first reference point, filter out the steady-state window before the event and the steady-state window after the event based on the first reference point, and extract steady-state features and electrical features based on the first reference point, the steady-state window before the event and the steady-state window after the event. The second feature extraction module is used to construct a transient window with the starting point of the steady-state window before the event as the starting point of the transient window and the ending point of the steady-state window after the event as the ending point of the transient window, and to extract transient features based on the transient window. The probability prediction module is used to visualize the steady-state features and the transient features respectively to obtain steady-state images and transient images, and input the steady-state images, the transient images, and the electrical features into a three-stream parallel deep learning network to obtain the probability of each identification result, and select the identification result corresponding to the maximum probability from all the probabilities of the identification results as the output result.

10. A storage medium, characterized in that, The storage medium stores one or more programs that, when executed by a processor, implement the non-intrusive load monitoring method based on a transient steady-state image as described in any one of claims 1-8.

Images