CN115423128A - Non-intrusive monitoring method for abnormal load behavior, electronic device and storage medium - Google Patents

Abstract

The invention relates to a non-invasive abnormal load behavior monitoring method, electronic equipment and a storage medium, wherein the method comprises the steps of obtaining real-time monitoring data of non-invasive load monitoring equipment and denoising the data; the monitoring data includes: total voltage and total current data of a preselected power circuit monitored by the non-intrusive load monitoring device; regarding the de-noised monitoring data, taking the monitoring data with load state conversion as effective monitoring data; based on a pre-constructed power strategy, carrying out color coding processing on effective monitoring data to obtain a V-I track image of total voltage and total current corresponding to each power circuit; and inputting the V-I track image into a training condition to generate a countermeasure network, and judging whether each power circuit has abnormal load or not based on the generated feature reconstruction image. The method has the beneficial effects that the technical problems of low expansibility, low flexibility and high monitoring error of non-intrusive load monitoring in the prior art can be solved.

Description

Non-intrusive monitoring method for abnormal load behavior, electronic device and storage medium

technical field

The invention relates to the technical field of load monitoring, in particular to a non-invasive monitoring method for abnormal load behavior, electronic equipment and a storage medium.

Background technique

In the production and life of modern society, a large amount of renewable energy is currently generated and utilized at the consumer end, and user behavior can promote the efficient integration of distributed energy that is highly dependent on weather. Therefore, it is crucial to observe the activities and actions of users using electricity. It is particularly important to obtain real-time power consumption information of various electrical appliances inside the user.

Different from the current smart meter and the acquisition of the total load power consumption information, the detailed monitoring of load power consumption is to obtain the real-time power consumption information of each electrical appliance in the power user through certain technical means, including the working status, power consumption and cumulative power of the electrical appliances. as well as fault information, etc.

In the existing technology, load monitoring mainly includes intrusive and non-intrusive. Intrusive load monitoring needs to intrude into the power load and install data measurement sensors with communication functions for each electrical appliance, and then collect and send electricity consumption information locally. . To achieve the same purpose, non-intrusive load monitoring only needs to install a data measurement sensor with communication function at the power supply entrance of the electric load, and can obtain the consumption of each electrical appliance in the user by analyzing the total load data. Telegram information. Compared to intrusive methods, non-intrusive methods are low cost, easy to install, and are well-applied with the detailed data obtained from non-intrusive load monitoring.

Due to the existence of fixed algorithm, the current non-intrusive load monitoring is only suitable for the ideal state where the type of power consumption load of power users is fixed and does not change. When the electrical equipment changes or ages, large monitoring errors will occur, and there Inflexible, non-scalable defects.

Contents of the invention

(1) Technical problems to be solved

In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a non-intrusive abnormal load behavior monitoring method, electronic equipment and storage media. The technical problems of low precision and high monitoring error.

(2) Technical solution

In order to achieve the above object, in the first aspect, the present invention provides a non-invasive method for monitoring abnormal load behavior, comprising the following steps:

S1. Obtain and denoise real-time monitoring data of the non-invasive load monitoring equipment; the monitoring data includes: total voltage and total current data of the preselected power circuit monitored by the non-invasive load monitoring equipment;

S2. For the denoised monitoring data, use the monitoring data with load state transition as effective monitoring data;

S3. Based on the pre-built power strategy, perform color-coding processing on the effective monitoring data, and obtain a V-I trajectory image of each single load in the power circuit;

S4. Input the V-I trajectory image into the trained conditional generative confrontation network, and judge whether there is an abnormal load in each power circuit based on the generated feature reconstruction image;

Wherein, the conditional generation confrontation network includes: a conditional autoencoder, a capsule network, and a classifier, the conditional autoencoder is used to convert the Gaussian prior probability of the load in the V-I trajectory image into a Gaussian posterior probability, and the capsule network is used for Achieving the compactness of the same class of features near the center of the Gaussian distribution makes the classifier detection load.

Optionally, before S1, the method also includes: S0, training the conditional generation confrontation network:

The S0 includes:

S01. Obtain training monitoring data samples and verification monitoring data samples used to train the condition generation confrontation network; the training monitoring data samples and the verification monitoring data samples are the historical monitoring total voltage and total current of the same power circuit data;

S02. Based on the pre-built power strategy, encode the training monitoring data samples and verification detection data samples, and acquire the training V-I trajectory image and verification V-I trajectory image of each single load in the power circuit;

S03. For each single load, input the training V-I trajectory image to the conditional confrontation generation network to reconstruct and generate a feature reconstruction image of the training V-I trajectory;

S04. Input the verification V-I trajectory image and the feature reconstruction image corresponding to each single load to a pre-built discriminator, and judge whether the feature reconstruction image matches the verification V-I trajectory image;

S05. Adjust the training parameters of the conditional confrontation generation network, and alternately generate the feature reconstruction image and input the discriminant network, so that the feature reconstruction image finally generated by the conditional confrontation monitoring network matches the verification V-I trajectory image. , to obtain the trained conditional adversarial monitoring network.

Optionally, S01 includes:

Obtain historical monitoring data of at least one load state transition event in the power circuit monitored by the non-intrusive load monitoring device; the load state transition event is the circuit load caused when a single load in the preselected power circuit is turned on and/or turned off conversion process;

Based on the pre-defined event detection window, calculate the time period when the load state transition event occurs; specifically:

Calculating the total real power S _t of the preselected power circuit, and determining the time t when ΔS _t >S _on1 ;

Based on the pre-built event detection window, calculate and determine when t=t+TR, the total real power variation ΔS _t+TR <S _on1 ; the R is the step size of the event detection window, ΔS _t =S _{t+ 1} -S _t ;

If S _t+TR -S _t <S _on2 , it is determined that a load state transition event occurs during the period t～t+TR;

Collecting the total voltage and total current data of T time periods before and after the occurrence of the load state transition event, and obtaining training monitoring data samples and verification monitoring data samples;

The S _on1 is a predefined load state transition event start threshold, and S _on2 is a predefined load state transition event end threshold.

Optionally, S3 includes:

S30. Based on a pre-built spectrum analysis method, sampling the effective monitoring data to obtain the voltage and current value of each single load in the power circuit;

S31. For each single load, based on the pre-built Fryze power strategy, determine the active component current i _a (t) and the reactive component current i _f (t) of the current i(t) of the single load;

所述功率因数为有功分量电流的功率与无功分量电流的功率的比值；Based on the active component current _ia ( _t ) and reactive component current if (t), calculate and obtain the power factor matrix

The power factor is the ratio of the power of the active component current to the power of the reactive component current;

的表达式为：The power factor matrix

The expression is:

K is the total number of sampling points, P _apparent is the real power, V _rms and I _rms are the effective values of the load voltage and current respectively;

和电压周期矩阵V；S32. For each single load, based on the pre-built HSV color space, construct the hue matrix of the VI track

and voltage periodic matrix V;

色相矩阵

和电压周期矩阵V，获取所述单个负荷的V-I轨迹图像。S33. For each single load, connect the power factor matrix in a standard three-dimensional coordinate system

Hue Matrix

and the voltage cycle matrix V to obtain the VI trajectory image of the single load.

Optionally, the S32 specifically includes;

S321. Based on the HSV color space, acquire the moving direction H _j of the VI trajectory by using the hue attribute hue;

Based on the motion direction H _j , store the hue of the jth sampling point in a 2N×2N matrix to obtain the hue matrix

The calculation expression of the direction of motion H _j is:

The arg is the arctangent function of four quadrants;

计算表达式为：The hue matrix

The calculation expression is:

|A|是集合的基数；

|A| is the cardinality of the set;

S322. Based on the pre-constructed binary image W _m (1,2,...,M), average M cycles of a single load voltage to obtain a voltage cycle matrix V;

The expression of the voltage cycle matrix V is:

Optionally, adjust the training parameters of the conditional confrontation monitoring network, specifically,

calculating a minimum value of the training loss of the conditional adversarial monitoring network based on a pre-built loss function;

adjusting parameters of the conditional generative adversarial network for the minimum weighted calculation;

The loss functions include a feature matching loss function, a reconstruction loss function, an additional encoder loss function, a center constraint loss function and/or a contrastive loss function.

Optionally, the feature matching loss function expression is:

The f(x) is the output of the middle layer of the discriminator for a given input V-I track x;

The expression of the reconstruction loss function is:

μ为训练V-I轨迹图像的平均强度，δ为训练V-I轨迹图像的标准差，

为训练V-I轨迹和特征重构图像的协方差；c1、c2为常数；said

μ is the average intensity of the training VI trajectory image, δ is the standard deviation of the training VI trajectory image,

Covariance of reconstructed images for training VI trajectory and features; c1 and c2 are constants;

The additional encoder loss function expression is:

为特征重构图像的编码特征；The z is the sampling vector feature output by the capsule network when training the VI trajectory image,

Reconstruct the encoded features of the image for the features;

The expression of the center constraint loss function is:

L _KL =d(C,sg[P _y ]);

所述C为概率胶囊，P为目标负荷簇的高斯分布；said

The C is a probability capsule, and P is a Gaussian distribution of target load clusters;

The expression of the contrast loss function is:

The [ ] ⁺ is a function that returns a positive number of parameters;

The formula for the weighted calculation of the minimum value is:

L=αL _KL +βL _rec +γL _contr +σL _enc +λL _adv , where α, β, γ, σ and λ are all constants.

Optionally, the S4 specifically includes:

The real-time V-I trajectory of the power circuit is input to the condition generation confrontation network of the training to generate a real-time feature reconstruction image;

Calculating the minimum distance between the real-time feature reconstruction image and the historical feature reconstruction image of the final training condition generation confrontation network;

If the minimum distance between the real-time feature reconstruction image and the historical feature reconstruction image is greater than the threshold τ satisfying the preset requirement, it is determined that an abnormal load occurs in the preselected power circuit.

In a second aspect, the present invention proposes an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor executes the computer program stored in the memory to realize any one of the above-mentioned first aspects. The steps of the method for the non-intrusive monitoring of abnormal load behavior.

In the third aspect, a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the non-intrusive abnormal load behavior described in any one of the above first aspects is realized The steps of the monitoring method.

(3) Beneficial effects

The present invention provides a non-intrusive monitoring method for abnormal load behavior, electronic equipment and storage media. In the method, the condition generation confrontation network is trained in advance, and the historical load voltage and Combined with the power strategy, the voltage and current values are color-coded to obtain the color load V-I trajectory image, which is conducive to visual recognition, and the V-I trajectory image and feature reconstruction image are input to the pre-built The conditional generative adversarial network, and the repeated discriminator judgment, realize the conditional generative adversarial network that meets the preset recognition correct ratio.

Compare the feature reconstruction image of the V-I trajectory image of the real-time collected voltage and current values with the minimum distance of the historical feature reconstruction image required by the load preset to judge whether there is an abnormal load, and realize the monitoring of the abnormal load of the circuit to be tested .

Compared with the prior art, the above-mentioned technical solution can realize the monitoring according to the actual load of the power user's load, achieve the purpose of flexible monitoring when the power user changes the electrical appliances, and improve the flexibility of non-intrusive abnormal load behavior monitoring, Scalability reduces detection errors.

Description of drawings

FIG. 1 is a schematic flowchart of a non-invasive method for monitoring abnormal load behavior provided by an embodiment of the present invention;

Fig. 2 is a training flowchart of the conditional generation confrontation network provided by an embodiment of the present invention;

FIG. 3 is a schematic diagram of a trained conditional generation confrontation network model provided by an embodiment of the present invention;

FIG. 4 is a schematic diagram of a logic flow for detecting a load state switching event provided by an embodiment of the present invention.

detailed description

In order to better explain the present invention and facilitate understanding, the present invention will be described in detail below through specific embodiments in conjunction with the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the invention may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be more clearly and thoroughly understood, and the scope of the present invention can be fully conveyed to those skilled in the art.

In today's society, electric energy has become one of the most important energy sources in modern society, and residents' demand for electricity is increasing day by day. Therefore, household energy management is a way to effectively reduce power waste. In the prior art, the non-invasive load monitoring method can only be better applied when the load type is fixed and does not change. For this reason, the present invention proposes a non-invasive load monitoring method. The intrusive monitoring method of abnormal load behavior can effectively respond to the load monitoring of various scenarios of adding or deleting electrical appliances or aging of electrical appliances in the household power circuit.

As shown in Figure 1, Figure 1 is a non-intrusive monitoring method for abnormal load behavior provided by an embodiment of the present invention. Non-intrusive load monitoring only needs to install a data storage device with communication function at the power supply entrance of the electric load. By using a measuring sensor, the power consumption information of each electrical appliance inside the user can be obtained by analyzing the total load data, and the method includes the following steps:

S1. Obtain and denoise real-time monitoring data of a non-intrusive load monitoring device; the monitoring data includes: total voltage and total current data of a preselected power circuit monitored by the non-intrusive load monitoring device.

S2. For the denoised monitoring data, use the monitoring data with load state transition as effective monitoring data.

S3. Based on the pre-built power strategy, perform color coding processing on the effective monitoring data, and obtain a V-I trajectory image of each single load in the power circuit.

In practical applications, the single load can be any electrical appliance in the household power circuit. Since the shape of the V-I trajectory depends largely on the load current reflecting the physical characteristics of the power load, the proportion of active current in the power factor is greater than the reactive current, resulting in a smaller difference in the shape of the V-I locus for loads of the same class. For this reason, in some embodiments, the Fryze power strategy can be used to decompose the load current into active current components representing resistive and non-resistive information and reactive current components, thereby enhancing the uniqueness of the V-I trajectory.

S4. Input the V-I trajectory image into the trained conditional generative adversarial network, and judge whether there is an abnormal load in each power circuit based on the generated feature reconstruction image.

Wherein, the conditional generation confrontation network includes: a conditional autoencoder, a capsule network, and a classifier, the conditional autoencoder is used to convert the Gaussian prior probability of the load in the V-I trajectory image into a Gaussian posterior probability, and the capsule network is used for Achieving the compactness of the same class of features near the center of the Gaussian distribution makes the classifier detection load.

A non-invasive monitoring method for abnormal load behavior proposed by the embodiment of the present invention obtains real-time monitoring data of non-invasive load monitoring equipment in real time and generates a V-I trajectory image, and inputs the V-I trajectory image into pre-trained conditions to generate In the confrontation network, it can judge whether there is an abnormal load and monitor the abnormal load. It is not affected by the change and aging of electrical equipment. It has good scalability, flexible use, and high applicability.

Specifically, the above non-invasive method for monitoring abnormal load behavior, S3 implemented in another embodiment may include:

S30. Based on a pre-built spectrum analysis method, sampling the effective monitoring data to obtain the voltage and current value of each single load in the power circuit;

S31. For each single load, based on the pre-built Fryze power strategy, determine the active component current i _a (t) and the reactive component current i _f (t) of the single load i(t);

Based on the active component current _ia ( _t ) and reactive component current if (t), calculate and obtain the power factor matrix

In this embodiment, the active current is defined as the orthogonal projection of the load current in the direction of the voltage v(t), that is, i _a (t) is proportional to v(t), which conveys resistance information, and the active current The expressions of i _a (t) and active power P _active are:

Among them, V _rms is the effective value of the voltage, and T is the power supply cycle. The reactive component current and voltage are orthogonal to each other, and the reactive current if ( _t ) can be represented by the instantaneous voltage and current of the load:

In practical applications, due to the large proportion of active current, the same type of load has little difference in the shape of the VI trajectory, so only the reactive current if (t) can be used to replace the current data without components i( _t ), based on the reactive current if ( _t ) to obtain the VI trajectory, while in order to avoid losing information between active and reactive components, saturation can be used to represent the ratio of active power to reactive power in multiple cycles, That is the power factor.

That is, the power factor is the ratio of the power of the active component current to the power of the reactive component current.

的表达式为：The power factor matrix

The expression is:

K is the total number of sampling points, P _apparent is the real power, V _rms and I _rms are the effective values of the load voltage and current, respectively.

和电压周期矩阵V。S32. For each single load, based on the pre-built HSV color space, construct the hue matrix of the VI track

and the voltage periodic matrix V.

During specific implementation, the S32 may include:

S321. For a single load, based on the HSV color space, use the hue attribute hue to obtain the moving direction H _j of the VI trajectory;

Based on the motion direction H _j , store the hue of the jth sampling point in a new 2N×2N matrix to obtain the hue matrix

The calculation expression of the direction of motion H _j is:

The arg is the arctangent function of the four quadrants; calculate the phase angle of two consecutive points in the V-I trajectory, and the value range is from 0° to 360°.

计算表达式为：The hue matrix

The calculation expression is:

|A|是集合的基数。

|A| is the cardinality of the set.

S322. Based on the pre-constructed binary image W _m (1,2,...,M), average the M cycles of the voltage to obtain the voltage cycle matrix V; that is, the color generation attribute Value using the HSV color space is used to represent the repeatability of the VI trajectory.

The expression of the voltage cycle matrix V is:

Wherein, m=1,2,3··M.

色相矩阵

和电压周期矩阵V，获取所述单个负荷的V-I轨迹图像。S33. For each single load, connect the power factor matrix in a standard three-dimensional coordinate system

Hue Matrix

and the voltage cycle matrix V to obtain the VI trajectory image of the single load.

In an embodiment, the value of M is preferably 10, which is determined according to actual conditions during application, and is not limited here.

The HSV (hue, saturation, lightness) color space is a non-linear conversion of RGB (red, green, blue color space), which is more in line with human perception of color. The HSV color space can be represented using an inverted cone model. The colors of each hue are distributed in radial slices from red to yellow, green, cyan, blue, and magenta. Hue is used to represent the category of colors. Saturation is defined as the ratio of hue to lightness and increases with distance from the center of a circular cross-section to indicate how vivid a color is. Luminosity represents brightness, and the distance from the center of the circle to the apex of the cone represents the lightness and darkness of each color.

色相矩阵

和电压周期矩阵V沿着第三维连接起来，将色相-饱和度-明度转换为等值的红-绿-蓝等以使创建的彩色图像可被人类感知。In an embodiment, the S33 is specifically implemented as, the power factor matrix

Hue Matrix

Concatenated with the voltage periodic matrix V along the third dimension, the hue-saturation-lightness is converted into equivalent red-green-blue etc. to create color images perceivable by humans.

Of course, in other embodiments, other converted colors may also be included, which is not limited here.

In some other embodiments, the S4 may specifically include:

The real-time V-I trajectory of the power circuit is input to the trained conditional generative confrontation network to generate real-time feature reconstruction images.

Calculating the minimum distance between the real-time feature reconstruction image and the historical feature reconstruction image of the finally trained conditional generative adversarial network.

If the minimum distance between the real-time feature reconstruction image and the historical feature reconstruction image is greater than the threshold τ satisfying the preset requirement, it is determined that an abnormal load occurs in the preselected power circuit.

In practical applications, the input load type can also be judged by the threshold, that is, if the minimum distance between the real-time feature reconstruction image and the historical feature reconstruction image is less than the threshold τ that meets the preset requirements, then the input load type can be judged The load type is the same as the load type corresponding to the minimum distance; if the minimum distance between the real-time feature reconstruction image and the historical feature reconstruction image is equal to the threshold τ that meets the preset requirements, it is determined that the input load type is an unknown load.

In some embodiments, it is determined that there is an abnormal load when the minimum distance between the real-time feature reconstructed image and the historical feature reconstructed image is greater than or equal to the threshold τ satisfying the preset requirement.

The above-mentioned embodiment of the present invention provides a non-intrusive monitoring method for abnormal load behavior, by separating the active component and reactive component of the load current, and only considering the reactive component as the original load current to generate V-I with a large difference Trajectory, based on this, the conditional generative adversarial network can better extract learning features and reduce recognition errors during training.

In some other embodiments, if the power circuit of the non-intrusive load monitoring device is used for the first time, before S1, the method may further include: S0, training the condition generation confrontation network, as shown in FIG. 2 , 2 is the training flowchart of the conditional generative adversarial network provided by an embodiment of the present invention.

The S0 may include:

S01. Obtain training monitoring data samples and verification monitoring data samples used to train the condition generation confrontation network; the training monitoring data samples and the verification monitoring data samples are the historical monitoring total voltage and total current of the same power circuit data;

Specifically, in some other embodiments, the training monitoring data samples and the verification monitoring data samples may be public data for non-intrusive load monitoring of pre-selected power circuits, and noise exists in the original data, which will affect the extraction of load characteristics , in order to facilitate subsequent feature extraction, denoising processing is usually performed on the data samples.

In a specific embodiment, said S01 is implemented as:

Obtain historical monitoring data of at least one load state transition event occurring in the power circuit monitored by the non-intrusive load monitoring device; the load state transition event is a circuit load transition caused when the device in the preselected power circuit is turned on and/or turned off process.

Based on the pre-defined event detection window, calculate the time period when the load state transition event occurs; specifically:

Calculating the total real power S _t of the preselected power circuit, and determining the time t when ΔS _t >S _on1 ;

Based on the pre-built event detection window, calculate and determine that when t=t+TR, the total real power variation ΔS _t+TR <S _on1 ; the R is the step size of the event detection window, ΔS _t =S _{t+ 1} -S _t ;

If S _t+TR −S _t <S _on2 , it is determined that a load state transition event occurs during the period t˜t+TR.

Collecting the total voltage and total current data of T periods of time before and after the occurrence of the load state transition event, and obtaining training monitoring data samples and verification detection data samples.

The S _on1 is a predefined load state transition event start threshold, and S _on2 is a predefined load state transition event end threshold.

S02. Based on the pre-built power strategy, encode the training monitoring data samples and verification detection data samples, and acquire the training V-I trajectory image and verification V-I trajectory image of each single load in the power circuit.

In actual operation, the process of generating the training V-I trajectory image and verifying the V-I trajectory image in S02 may be the same as the sub-steps of generating the real-time V-I trajectory image in the above embodiment.

S03. For each single load, input the training V-I trajectory image to the conditional confrontation generation network to reconstruct and generate a feature reconstruction image of the training V-I trajectory.

S04. Input the verification V-I trajectory image and the feature reconstruction image corresponding to each single load to a pre-built discriminator, and judge whether the feature reconstruction image matches the verification V-I trajectory image.

S05. Adjust the training parameters of the conditional confrontation generation network, and alternately generate the feature reconstruction image and input the discriminant network, so that the feature reconstruction image finally generated by the conditional confrontation monitoring network matches the verification V-I trajectory image. , to obtain the trained conditional adversarial monitoring network.

In some embodiments, in S05, adjusting the training parameters of the conditional confrontation monitoring network can be specifically implemented as:

calculating a minimum value of the training loss of the conditional adversarial monitoring network based on a pre-built loss function;

adjusting parameters of the conditional generative adversarial network for the minimum weighted calculation;

In a specific embodiment, the loss function may include a feature matching loss function, a reconstruction loss function, an additional encoder loss function, a center constraint loss function and/or a contrast loss function, and the like.

Specifically, the feature matching loss function is used for confrontation learning to reduce the instability of conditional generation confrontation network training, and the conditional generation confrontation network encodes the generated feature distribution and the generated VI trajectory feature reconstruction image with the real VI trajectory Image alignment, based on the feature matching loss function, generates VI trajectories sufficient to fool the discriminator, which can effectively distinguish feature representations of known devices from unknown devices. Specifically, the generator is updated based on the internal representation of the discriminator. Formally, let f be a function of a given input VI trajectory x output discriminator intermediate layer drawn according to the input data distribution, feature matching respectively computes the _L2 distance between the feature representation of the original VI trajectory image and the generated VI trajectory image .

The feature matching loss function, that is, the expression of the adversarial loss function is:

The f(x) is the output of the middle layer of the discriminator according to the input V-I track x.

In another embodiment, the reconstruction loss between the input and the reconstructed V-I trajectory image can also be measured to solve the unoptimized use of the context information of the input V-I data to obtain a credible reconstruction result, based on the V-I trajectory image The generation process contains rich structural information. In some embodiments, structural similarity loss is used as the reconstruction loss of the generator. The structural similarity loss takes into account brightness, contrast and structural information, and is not sensitive to the position offset of the input V-I trajectory and its reconstruction is too sensitive, thus making it easier for the network to converge. Therefore, models trained with structural similarity loss tend to focus on global information rather than local features during V-I trajectory reconstruction.

The expression of the reconstruction loss function, that is, the structural similarity loss function is:

μ为训练V-I轨迹图像的平均强度，δ为训练V-I轨迹图像的标准差，

为训练V-I轨迹和特征重构图像的协方差；c1、c2为常数，在一些实施例中，所述常数c1和c2分别设置为0.01和0.03。said

μ is the average intensity of the training VI trajectory image, δ is the standard deviation of the training VI trajectory image,

Covariance of reconstructed images for training VI trajectory and features; c1 and c2 are constants, and in some embodiments, the constants c1 and c2 are set to 0.01 and 0.03, respectively.

Based on the above two loss functions, the generator can be forced to produce images that are real and contextual.

之间的距离，基于此，条件生成对抗网络学习如何对已知负荷样本的V-I轨迹特征进行编码，生成器和额外的编码器网络都只针对已知负荷的数据样本进行优化。Further, in some other embodiments, an additional encoder loss L ₁ can also be used to minimize the sampling vector features of the capsule network output from the input z and the encoded features of the reconstructed VI trajectory image

Based on the distance between , the conditional generative adversarial network learns how to encode the VI trajectory features of the known load samples, and both the generator and the additional encoder network are optimized only for the known load data samples.

The additional encoder loss function expression is:

为特征重构图像的编码特征。The z is the sampling vector feature output by the capsule network when training the VI trajectory image,

Reconstruct the encoded features of an image for features.

In some other embodiments, in order to encode the VI trajectory features of each type of known power circuit load, each type of load feature forms a compact cluster, making it easier for the model to identify unknown load features. A center-constrained loss can also be used in the latent space of the generator to push the probability capsule C towards the center of the target loading cluster P _y , so that the density of all known loading samples is concentrated in the target area.

The expression of the center constraint loss function is:

L _KL =d(C,sg[P _y ]);

The function sg[ ] represents the stopping gradient operator, which is defined as the identity in the forward computation and has zero partial derivatives, restricting its parameters to non-updated constants.

所述C为概率胶囊，P为目标负荷簇的高斯分布；所述概率胶囊为输入V-I轨迹的高斯分布。said

The C is a probability capsule, and P is the Gaussian distribution of the target load cluster; the probability capsule is the Gaussian distribution of the input VI trajectory.

In one embodiment, a comparative loss function is also constructed, using the margin Loss boundary loss function and the marginm _k boundary to push all target loads that do not belong to y far away from the distribution C, by considering that P _{≠ y} is the difference of P _y , Collapse of the conditional generative adversarial network's previous target load is avoided, facilitating the separation between the load and all other loads (possibly unknown corresponding loads).

The expression of the contrast loss function is:

The [·] ⁺ is a function that returns a positive number of parameters.

Based on the above five loss functions, the formula for updating the parameters of the network by calculating the weighted combination of the minimum values of the five loss functions is:

L = αL _KL + βL _rec + γL _contr + σL _enc + λL _adv .

The α, β, γ, σ and λ are all constants. In one embodiment, α is preferably 1, β is preferably 0.01, γ is preferably 1, σ is preferably 0.01, and λ is preferably 10.

In this embodiment, the conditional autoencoder and the capsule network are used as the generator of the conditional generative adversarial network. In the process of model training, the capsule features of the same load are matched with a predetermined Gaussian distribution. For each class The load defines a Gaussian distribution, specifically, using a variational autoencoder framework, using a set of Gaussian priors as an approximation of the posterior distribution, so that the compactness of the same class of features near the center of the Gaussian distribution can be controlled, thereby controlling the classification The ability of the loader to detect unknown loads. The V-I trajectory generated by the generator passes through the discriminator and an additional encoding network, and the additional encoding network maps the generated V-I trajectory to the hidden layer representation to minimize the distance between the hidden layer representation of the V-I trajectory in the generator, through Construct multiple loss functions and find the minimum value to adjust the model parameters, so that the generator can better learn the distribution characteristics of known loads, increase its ability to monitor unknown loads, and have high monitoring flexibility and small errors.

As shown in Figure 3, Figure 3 is a schematic diagram of the model of the trained conditional generation confrontation network provided by an embodiment of the present invention. In the embodiment shown in Figure 3, the conditional generation confrontation network includes a conditional autoencoder and a capsule network , also includes an additional encoder. Each loading has a non-individual Gaussian distribution, and the conditional autoencoder approximates the Gaussian prior of the loading as the posterior probability. The additional encoder is used to map the generated V-I trajectory to the hidden layer representation to minimize the distance from the hidden layer representation of the V-I trajectory in the conditional autoencoder. The capsule network is used to achieve the compactness of the same class of features around the center of the Gaussian distribution, thereby controlling the ability of the classifier to detect unknown loads.

In order to better explain the technical solution proposed by the present invention, a detailed description will be made below in conjunction with a specific embodiment.

This embodiment is a monitoring of abnormal load behavior of a household power circuit using a non-intrusive load monitoring device. In the household power circuit, each electrical appliance/equipment/appliance is regarded as a single load. The load state switching process is accompanied by changes in real power. Motor loads are often accompanied by changes in power and current effective values when starting, and load state changes will occur. This process is regarded as a load state switching event. The variation of the power or current RMS can be compared with a preset threshold, and if it is greater than the threshold, it is judged that an event has occurred. According to the voltage and current changes before and after the event, the voltage and current values of the load that caused the event can be obtained.

First, a conditional generative network is trained; the network is trained using only V-I trajectories with known loads.

Steps include:

A1. Obtain the original monitoring data of the non-invasive load monitoring equipment, and denoise the original monitoring data;

A2. Based on the predefined event detection window, calculate the time period during which the load state transition event occurs.

As described in FIG. 4 , FIG. 4 is a schematic diagram of a logic flow for detecting a load state switching event provided in this embodiment.

In this embodiment, the detection load state switching event is specifically implemented as follows:

The step size of the time detection window is defined as R, S _t represents the total real power at t seconds, and ΔS _t =S _t+1 −S _t represents the total real power variation. When ΔS _t >S _on1 , the event detection window starts to move and calculate ΔS _t+1 , ΔS _t+2 . . . until ΔS _t+TR <S _o1n . If S _t+TR −S _t <S _o2n , it means that a load has a state change within t˜t+TR seconds, that is, an unknown load state switching event is detected. The load state switching event start time t _on is t seconds, the event end time t _off is t+TR, and TR represents the duration of the event.

The detection process of the above load state transition event can be expressed by the following formula:

ΔS _t |≥S _on1 &&|ΔS _t+1 |≥S _on1 &&...&&|ΔS _t+TR-1 |≥S _on1

&&|ΔS _t+TR |<S _on1 &&|ΔS _t+TR+1 |<S _on1 &&|S _t+TR −S _t |≥S _on2 .

A3. Extract the steady-state voltage and current waveforms of T periods before and after the load state switching event, and obtain the voltage and current value of a single load based on the spectrum analysis method;

Specifically, the steady-state voltage and current waveforms v _on , v _off , i _on , i _off of T periods before and after the load state transition event are extracted, and the base voltage phase angle is calculated by fast Fourier transform and other spectral analysis methods, and then the phase angle The zero sampling point is used as the initial sampling point to ensure that the current waveforms i _off and i _on can be directly subtracted in the time domain. The voltage of a single load is v=(v _off +v _on )/2 and the current i=i _{off −ion} .

V _on refers to the total voltage after the load state switching event, V _off refers to the total voltage before the load state switching event, I _on refers to the total current after the load state switching event, and I _off refers to the total current before the load state switching event.

A4. Use the Fryze power strategy to color-code the voltage and current values of the above-mentioned single load, and obtain the load V-I trajectory image of the load in color. The V-I trajectory image is divided into a training V-I trajectory image and a verification V-I trajectory image, and the conditional generation confrontation network is trained until the correct recognition rate of the conditional generation confrontation network is 95%.

The correct recognition rate of the conditional generation confrontation network is 95%, and the threshold τ for identifying abnormal loads is set. The threshold τ is the distribution threshold of the Gaussian distribution of the V-I trajectory image reconstructed by the conditional generation confrontation network for the same load.

In this embodiment, according to the actual needs of this embodiment, the correct recognition rate of the trained conditional generative adversarial network is 95%. In other embodiments, it is confirmed according to actual needs, which is not limited here.

Then, the trained conditional confrontation generation network is used to monitor the power circuit to determine whether there is an abnormal load.

Specifically include:

B1. Obtain real-time monitoring data of the power circuit, and denoise, the real-time monitoring data is the total voltage and total current of the power circuit;

B2. For the monitoring data after denoising, use the monitoring data of load state transition as effective monitoring data; in this embodiment, the process of judging the monitoring data of load state transition is the same as that in A2, and the use of phase synchronization Long event detection window.

B3. Extract the steady-state voltage and current waveforms of T periods before and after the load state switching event, and obtain the voltage and current value of a single load based on the spectrum analysis method; the step flow is the same as the above-mentioned A3, and is calculated by using spectrum analysis methods such as fast Fourier transform The base voltage phase angle, and then the sampling point where the phase angle is zero is taken as the initial sampling point to ensure that the current waveforms i _off and i _on can be directly subtracted in the time domain.

B4. Use the Fryze power strategy to color-code the voltage and current values of the above-mentioned single load, and obtain the load V-I trajectory image of the load in color.

B5. Input the V-I trajectory image into the trained conditional generative adversarial network, and judge whether there is an abnormal load in each power circuit based on the generated feature reconstruction image.

Input the V-I trajectory to be tested into the trained conditional generation confrontation network. If the minimum distance between the capsule feature of the V-I trajectory of the load, that is, the feature reconstruction map and the Gaussian distribution of each known load is greater than or equal to the threshold, it is determined that there is Abnormal load occurs. If the minimum distance to the Gaussian distribution of the known load is less than the threshold, then the load category label is the load category label corresponding to the minimum distance.

In addition, the present invention proposes an electronic device, including a memory and a processor, the memory stores a computer program, and the processor executes the computer program stored in the memory to realize the non-invasive Steps in a method for monitoring abnormal load behavior.

In practical applications, human-computer interaction equipment can also be set to enable users to view the current monitoring results in real time.

The present invention also provides a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the non-intrusive abnormal load behavior as described in any one of the above-mentioned embodiments is realized The steps of the monitoring method.

The invention provides a non-intrusive monitoring method for abnormal load behavior, electronic equipment and a storage medium, through training conditions to generate an adversarial network, so as to realize the monitoring of abnormal load behavior and the correct identification of known loads.

The generator of the conditional generation confrontation network provided by an embodiment of the present invention is composed of an autoencoder and a capsule network. The capsule features of the V-I trajectory of the same load after passing through the encoder and the capsule network correspond to a predefined Gaussian distribution. Each load has A uniquely Gaussian distribution, the conditional autoencoder approximates the Gaussian prior of the loading to the posterior probability. An additional encoding network maps the V-I trajectory generated by the generator to the hidden representation to minimize the distance from the hidden representation of the V-I trajectory in the generator. By constructing multiple loss functions to minimize each loss, adjust and update the parameters of the conditional generation confrontation network, and realize the compactness of the same load feature near the center of the Gaussian distribution. The conditional generation confrontation network learns how to encode the V-I trajectory characteristics of the load, and Learning the data distribution of known loads can realize real-time monitoring of unknown loads or abnormal loads among power users, with high scalability, strong flexibility, and low error.

The non-intrusive abnormal load behavior monitoring method, electronic equipment, and storage medium provided by each embodiment of the present invention have high accuracy in monitoring abnormal load behavior, and are applied in power circuit systems such as households, and can obtain real-time data of various electrical appliances inside the user. Power consumption information, including the working status of electrical appliances, power consumption, accumulated power, and fault information. It is conducive to the formulation of energy efficiency policies, and avoids the occurrence of the aging of the circuit hardware inside the load, which will cause the electrical appliances to fail and fail to work normally, and even bring direct economic losses to power users. Moreover, the method provided by the invention has low installation cost, flexible and convenient application, high expansibility and good application prospect.

In the description of the present invention, it should be understood that the terms "first" and "second" are used for description purposes only, and cannot be interpreted as indicating or implying relative importance or implicitly indicating the quantity of indicated technical features. Thus, a feature defined as "first" and "second" may explicitly or implicitly include one or more of these features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

In the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" refers to the A particular feature, structure, material, or characteristic is described as included in at least one embodiment or example of the invention. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Furthermore, the described specific features, structures, materials or characteristics may be combined in any suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples and features of different embodiments or examples described in this specification without conflicting with each other.

Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and those skilled in the art can make the above-mentioned The embodiments are subject to alterations, modifications, substitutions and variations.

Claims (10)

1. A monitoring method of non-invasive abnormal load behavior, characterized in that, comprising the following steps: S1. Obtain and denoise real-time monitoring data of the non-invasive load monitoring equipment; the monitoring data includes: total voltage and total current data of the preselected power circuit monitored by the non-invasive load monitoring equipment; S2. For the denoised monitoring data, use the monitoring data with load state transition as effective monitoring data; S3. Based on the pre-built power strategy, perform color-coding processing on the effective monitoring data, and obtain a V-I trajectory image of each single load in the power circuit; S4. Input the V-I trajectory image into the trained conditional generative confrontation network, and judge whether there is an abnormal load in each power circuit based on the generated feature reconstruction image; Wherein, the conditional generation confrontation network includes: a conditional autoencoder, a capsule network, and a classifier, the conditional autoencoder is used to convert the Gaussian prior probability of the load in the V-I trajectory image into a Gaussian posterior probability, and the capsule network is used for Achieving the compactness of the same class of features near the center of the Gaussian distribution makes the classifier detection load. 2. The monitoring method according to claim 1, characterized in that: Before S1, the method also includes: S0, training the conditional generation confrontation network: The S0 includes: S01. Obtain training monitoring data samples and verification monitoring data samples used to train the condition generation confrontation network; the training monitoring data samples and the verification monitoring data samples are the historical monitoring total voltage and total current of the same power circuit data; S02. Based on the pre-built power strategy, encode the training monitoring data samples and verification detection data samples, and acquire the training V-I trajectory image and verification V-I trajectory image of each single load in the power circuit; S03. For each single load, input the training V-I trajectory image to the conditional confrontation generation network to reconstruct and generate a feature reconstruction image of the training V-I trajectory; S04. Input the verification V-I trajectory image and the feature reconstruction image corresponding to each single load to a pre-built discriminator, and judge whether the feature reconstruction image matches the verification V-I trajectory image; S05. Adjust the training parameters of the conditional confrontation generation network, and alternately generate the feature reconstruction image and input the discriminant network, so that the feature reconstruction image finally generated by the conditional confrontation monitoring network matches the verification V-I trajectory image. , to obtain the trained conditional adversarial monitoring network. 3. monitoring method as claimed in claim 2 is characterized in that, S01 comprises: Obtain historical monitoring data of at least one load state transition event in the power circuit monitored by the non-intrusive load monitoring device; the load state transition event is the circuit load caused when a single load in the preselected power circuit is turned on and/or turned off conversion process; Based on the pre-defined event detection window, calculate the time period when the load state transition event occurs; specifically: Calculating the total real power S _t of the preselected power circuit, and determining the time t when ΔS _t >S _on1 ; Based on the pre-built event detection window, calculate and determine that when t=t+TR, the total real power variation ΔS _t+TR <S _on1 ; the R is the step size of the event detection window, ΔS _t =S _{t+ 1} -S _t ; If S _t+TR -S _t <S _on2 , it is determined that a load state transition event occurs during the period t～t+TR; Collecting the total voltage and total current data of T time periods before and after the occurrence of the load state transition event, and obtaining training monitoring data samples and verification monitoring data samples; The S _on1 is a predefined load state transition event start threshold, and S _on2 is a predefined load state transition event end threshold. 4. The monitoring method according to claim 1, wherein S3 comprises: S30. Based on a pre-built spectrum analysis method, sampling the effective monitoring data to obtain the voltage and current value of each single load in the power circuit; S31. For each single load, based on the pre-built Fryze power strategy, determine the active component current i _a (t) and the reactive component current i _f (t) of the current i(t) of the single load; Based on the active component current _ia ( _t ) and reactive component current if (t), calculate and obtain the power factor matrix

The power factor is the ratio of the power of the active component current to the power of the reactive component current; The power factor matrix

The expression is:

K is the total number of sampling points, P _apparent is the real power, V _rms and I _rms are the effective values of the load voltage and current respectively; S32. For each single load, based on the pre-built HSV color space, construct the hue matrix of the VI track

and voltage periodic matrix V; S33. For each single load, connect the power factor matrix in a standard three-dimensional coordinate system

Hue Matrix

and the voltage cycle matrix V to obtain the VI trajectory image of the single load. 5. monitoring method as claimed in claim 4, is characterized in that, The S32 specifically includes; S321. Based on the HSV color space, acquire the moving direction H _j of the VI trajectory by using the hue attribute hue; Based on the motion direction H _j , store the hue of the jth sampling point in a 2N×2N matrix to obtain the hue matrix

The calculation expression of the direction of motion H _j is:

The arg is the arctangent function of four quadrants; The hue matrix

The calculation expression is:

|A| is the cardinality of the set; S322. Based on the pre-constructed binary image W _m (1,2,...,M), average M cycles of a single load voltage to obtain a voltage cycle matrix V; The expression of the voltage cycle matrix V is:

6. The monitoring method according to claim 2, wherein adjusting the training parameters of the conditional confrontation monitoring network is specifically, calculating a minimum value of the training loss of the conditional adversarial monitoring network based on a pre-built loss function; adjusting parameters of the conditional generative adversarial network for the minimum weighted calculation; The loss functions include a feature matching loss function, a reconstruction loss function, an additional encoder loss function, a center constraint loss function and/or a contrastive loss function. 7. monitoring method as claimed in claim 6 is characterized in that, The expression of the feature matching loss function is:

The f(x) is the output of the middle layer of the discriminator for a given input V-I track x; The expression of the reconstruction loss function is:

said

μ is the average intensity of the training VI trajectory image, δ is the standard deviation of the training VI trajectory image,

Covariance of reconstructed images for training VI trajectory and features; c1 and c2 are constants; The additional encoder loss function expression is:

The z is the sampling vector feature output by the capsule network when training the VI trajectory image,

Reconstruct the encoded features of the image for the features; The expression of the center constraint loss function is: L _KL =d(C,sg[P _y ]); said

The C is a probability capsule, and P is a Gaussian distribution of target load clusters; The expression of the contrast loss function is:

The [·] ⁺ is a function that returns a positive number of parameters; The formula for the weighted calculation of the minimum value is: L=αL _KL +βL _rec +γL _contr +σL _enc +λL _adv , where α, β, γ, σ and λ are all constants. 8. monitoring method as claimed in claim 1, is characterized in that, The S4 specifically includes: The real-time V-I trajectory of the power circuit is input to the condition generation confrontation network of the training to generate a real-time feature reconstruction image; Calculating the minimum distance between the real-time feature reconstruction image and the historical feature reconstruction image of the final training condition generation confrontation network; If the minimum distance between the real-time feature reconstruction image and the historical feature reconstruction image is greater than the threshold τ satisfying the preset requirement, it is determined that an abnormal load occurs in the preselected power circuit. 9. An electronic device, characterized in that it comprises a memory and a processor, wherein a computer program is stored in the memory, and the processor executes the computer program stored in the memory to realize any of the above claims 1 to 8. The steps of the non-intrusive abnormal load behavior monitoring method described above. 10. A computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the non-intrusive device according to any one of claims 1 to 8 is realized. Steps in a method for monitoring abnormal load behavior.

Images