A non-intrusive load identification method based on multimodal features

Abstract

The present invention proposes a non-intrusive load identification method based on multimodal features, including the following steps: 1) load data acquisition; 2) event detection based on power or current effective value, extracting the transient section of the event and the steady-state section before and after the transient section of the event; 3) based on the target indicator sequence of the transient section of the event, extracting the path signature feature of the target indicator sequence, if the collected load data includes high-frequency voltage data and high-frequency current data, then go to step 4, otherwise go to step 5; 4) based on the high-frequency voltage and high-frequency current of the steady-state section before and after the event, extract the VI trajectory image feature; 5) based on the data condition selection extraction corresponding load imprint, use the twin network framework for load identification. The present invention extracts the transient path signature features of low-frequency data and the VI trajectory features of high-frequency data according to the collected data frequency, and realizes load identification through the mid-end fusion method of multimodal features and the twin network.

Description

Technical Field

[0001] The invention relates to a non-intrusive load monitoring method. Technical Background

[0002] The load monitoring system is an important part of residents' demand response and safe electricity use. The current load monitoring systems are roughly divided into two categories: invasive and non-invasive. The traditional invasive load monitoring system installs sensors at each load to monitor the operation of each load. Due to its high hardware economic cost, low system reliability, and small user coverage, it is difficult to be widely used. The non-intrusive load monitoring (NILM) system is characterized by installing monitoring equipment at the power entrance, which can analyze the various power loads within the user and provide information such as the switching and energy consumption of each power device.

[0003] The transformation process of the working state of electrical equipment is called transient event. Transient characteristics refer to the unique transient characteristic information during transient events, such as the current, voltage, and power change information at the moment of the event. Steady-state characteristics refer to the voltage, current, power, harmonics, and other information before and after the event. The unique characteristic information of electrical equipment when its state changes is called load imprint.

[0004] The construction of load imprint is the key to the accuracy of non-intrusive load identification. Existing load imprint construction methods include power change values ​​based on low-frequency data, transient power curves based on low-frequency data, VI trajectory characteristics based on high-frequency steady-state data, etc. Transient characteristics and steady-state characteristics can often only represent part of the load characteristics. Common load identification models include K-nearest neighbor algorithm, hidden Markov algorithm, convolutional neural network, etc. The load identification accuracy based on hidden Markov algorithm is generally low. The supervision model based on low-frequency load imprint is difficult to distinguish power-similar loads. The identification method based on VI trajectory and neural network has a higher accuracy, but the required high-frequency data has high requirements on hardware facilities and requires more labeled data for model training. Summary of the invention

[0005] The present invention proposes a non-intrusive load identification method based on multimodal features, aiming to solve at least one of the technical problems existing in the above-mentioned prior art.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] A non-intrusive load identification method based on multimodal features, using load data collected at the power inlet, includes the following steps:

[0008] Step 1: Obtain load data. Obtain the collected load data, assuming that the frequency of the load data is at least greater than 50 Hz. If the frequency is greater than 1 kHz, the load data includes high-frequency voltage data, high-frequency current data, active power P, reactive power Q, and effective current value I; otherwise, the load data includes active power P, reactive power Q, and effective current value I.

[0009] Step 2: Perform event detection based on the power or current effective value, and extract the event transient segment and the steady-state segments before and after the event transient segment.

[0010] Using the current effective value, event detection is performed based on the composite sliding window and bilateral correction method to extract the start and end points of the transient process. The specific process is as follows:

[0011] Set the statistical sliding window W s and the detection sliding window W d , where W s Length L s , W d Length L d , I i-Ls-1 and I i-1 are the starting point and end point of the current statistical sliding window respectively, calculate the mean value μ and standard deviation σ of the current effective value of the statistical sliding window, and calculate the detection sliding window W d The change statistic Z-score on:

[0012]

[0013] Use the sliding window to traverse the current effective value sequence and find that Z is less than the threshold Z threshold The point values ​​are set to zero. After obtaining the Z value sequence, the sequence is divided into N categories through density clustering, where the corresponding point sequence in each category is a transient event process T.

[0014] For the transient event process T obtained, bilateral correction processing is performed by calculating the mean drift distance and slope. The point with the largest mean drift distance near the starting point of the transient event is corrected as the starting point, and the point where the slope is less than the slope threshold for the first time before the end point of the transient event is corrected as the end point, so as to obtain the accurate transient event starting point I start With stop point I end .

[0015] Starting Point I start With stop point I end The transient section of the event is T, starting point I start O before n The period constitutes the pre-event steady-state period O b , stop point I end After O n The period constitutes the post-event steady-state period O a .

[0016] Step 3: Based on the target indicator sequence S of the event transient section, extract the path signature characteristics of the target indicator sequence S. If the collected load data includes high-frequency voltage data and high-frequency current data, go to step 4, otherwise go to step 5.

[0017] In order to obtain features suitable for deep learning model input representation, the path signature features of the target indicator sequence S are extracted. The path signature is the calculation result of the rough path theory.

[0018] For a scalar field F: Then for the path The path integral of is:

[0019]

[0020] where r:[a,b]→X is the parametric equation of the path.

[0021] For n-th path signature items, there are:

[0022]

[0023] Path X:[a,b]→R n The path signature S(X) a,b , is an infinitely long sequence consisting of signature terms of various orders:

[0024]

[0025] Set the active power P, reactive power Q or current effective value I of the transient section to form the target indicator sequence S, and extract the n-order truncated path signature S of the sequence S T Used as features for machine learning input.

[0026] Step 4: Extract VI trajectory image features based on the high-frequency voltage and high-frequency current in the steady-state section before and after the event:

[0027] Extract the steady-state segment before the event O b The high-frequency voltage and current data are recorded as v b andi b .

[0028] Extract the post-event steady-state segment O a The high-frequency voltage and current data are recorded as v a andi a .

[0029] v b 、i b 、v a andi aSmoothing is performed to replace outliers with smoothed values, thereby avoiding the impact of outliers.

[0030] v b 、i b 、v a andi a Perform three Hermite interpolation fits respectively to obtain the interpolation function f vb 、f ib 、f va and f ia .

[0031] v b 、i b 、v a andi a Perform fast Fourier transform respectively to get the fundamental wave phase angle and And calculate the displacement of the waveform when the fundamental phase angle is zero, and get the initial sampling time t vb ,t ib ,t va and t ia , in the interpolation function f vb 、f ib 、f va 、f ia t vb ,t ib ,t va ,t ia Resample the same number of samples f at the initial sampling time, where f is O n To ensure that the sampling points of each cycle are aligned, the phase angle of v′ is aligned after resampling. b , i′ b , v′ a and i′ a .

[0032] v′ b , i′ b , v′ a and i′ a Medium n The average value of the sampling points is calculated for each cycle, and v is obtained. avr b 、i avr b 、v avr a 、i avr b .

[0033] Generate voltage sequence v ob and current sequence i ob :

[0034]

[0035] i ob =i avr a -i avr b

[0036] With voltage sequence v ob As the horizontal axis, the current sequence i ob Generate a VI trajectory graph for the vertical axis. The process is as follows:

[0037] Set the image pixel matrix A T Size (M, M), the initial pixel value is (255,255,255);

[0038] v ob 、i ob Resampled to M points, the pixel position of the jth voltage and current (x j ,y j )for:

[0039]

[0040]

[0041] Where int is the rounding function.

[0042] make

[0043]

[0044] If the jth voltage and current are in an upward trend in the voltage cycle, the pixel position (x j ,y j ) is taken as (value,value,255). If the jth voltage and current are in a downward trend in the voltage cycle, the pixel position (x j ,y j ) is taken as (value,value,value). The extracted image features are recorded as A T .

[0045] Step 5: Select and extract the corresponding load imprint based on data conditions, and use the twin network framework for load identification.

[0046] If the collected load data includes high-frequency voltage and current data, the path signature feature S is selected. T and VI trajectory image feature A T As a load imprint, build a load imprint library of target device events.

[0047] If the collected load data does not include high-frequency voltage and current data, the path signature feature S is selected. T As a load imprint, build a load imprint library of target device events.

[0048] If the collected load data includes high-frequency voltage and current data, the path signature feature S is selected. T and VI trajectory image feature A T As the load mark, denoted as F T =[S T , A T ].

[0049] If the collected load data does not include high-frequency voltage and current data, the path signature feature S is selected. T As the load mark, denoted as F T =S T .

[0050] Perform pairwise transformation on features. If T i With T j Belong to the same type of event, then the label y ij is 1, otherwise the label y ij is 0, and the sample form is as follows in For input 1, For input 2, y ij is the output.

[0051] The twin neural network has two inputs. The two inputs are sent to two neural networks. The two neural networks with shared weights map the inputs to new spaces respectively, forming representation vectors of the inputs in the new space. The similarity of the two inputs is evaluated as output through the calculation of the loss function. If the collected load data does not include high-frequency voltage and current data, the path signature feature S is converted into T Form a representation vector in the new space; if the collected load data includes high-frequency voltage and current data, the path signature feature S T and VI trajectory image feature A T The mid-end fusion of multimodal data is performed through the fully connected layer and the convolutional layer to form a representation vector.

[0052] The trained twin network model is used for load identification. The process is as follows:

[0053] Takes as input the payload imprint of the detected unknown device event The load imprint library has K load imprints. The load imprints in the load imprint library are traversed and the load imprints in the load imprint library are used as input in sequence. Use the trained twin network model to get the output y ij, so there are K outputs y for the load imprint of the unknown device event ij , if the output y ij If the maximum value is greater than the threshold, the output y ij The maximum value corresponds to The event of the corresponding device is used as the load identification result, otherwise the event type is identified as an unknown event.

[0054] The implementation of the present invention has the following beneficial effects:

[0055] The present invention can automatically select the corresponding load imprint based on data conditions. The transient path signature features can be extracted based on low-frequency data, which can effectively represent the path trajectory features of the transient process. The VI trajectory features are extracted based on high-frequency data, which can effectively represent the steady-state voltage and current trajectory features before and after the event. The path signature and VI trajectory fusion features effectively and comprehensively represent the transient and steady-state characteristics of the event. The twin network is a method of few-sample learning. For each type of equipment, only a small number of samples need to be labeled to achieve good results. The present invention reduces the requirements for data conditions in two dimensions: acquisition frequency and sample labeling. It is a practical and effective load identification method. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is an overall schematic diagram of a non-intrusive load identification method based on multimodal features of the present invention.

[0057] Figure 2 It is a schematic diagram of a transient section and a steady-state section of an embodiment of the present invention.

[0058] Figure 3 It is a VI trajectory diagram of a microwave oven start-up event according to an embodiment of the present invention.

[0059] Figure 4 It is a twin network framework diagram of an embodiment of the present invention.

[0060] Figure 5 It is a network structure diagram of an embodiment of the present invention using path signature features as input.

[0061] Figure 6 It is a network structure diagram of an embodiment of the present invention using multimodal features as input. DETAILED DESCRIPTION

[0062] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0063] Figure 1 The present invention shows an execution flow of a non-intrusive load identification method based on multimodal features, which includes the following steps:

[0064] Step 1: load data acquisition;

[0065] Step 2: Perform event detection based on power or current effective value, extract the event transient section and the steady-state sections before and after the event transient section;

[0066] Step 3: based on the target indicator sequence of the transient section of the event, extract the path signature features of the target indicator sequence. If the collected load data includes high-frequency voltage data and high-frequency current data, go to step 4, otherwise go to step 5.

[0067] Step 4, extracting VI trajectory image features based on the high-frequency voltage and high-frequency current in the steady-state section before and after the event;

[0068] Step five: Select and extract the corresponding load imprint based on data conditions, and use the twin network framework for load identification.

[0069] The above steps are described in detail as follows:

[0070] Step 1: Obtain load data. Obtain the collected load data, assuming that the frequency of the load data is at least greater than 50 Hz. If the frequency is greater than 1 kHz, the load data includes high-frequency voltage data, high-frequency current data, active power P, reactive power Q, and effective current value I; otherwise, the load data includes active power P, reactive power Q, and effective current value I.

[0071] Step 2: Perform event detection based on the power or current effective value, and extract the event transient segment and the steady-state segments before and after the event transient segment.

[0072] In the embodiment of the present invention, the current effective value is used to detect events based on a composite sliding window and bilateral correction method to extract the start and end points of the transient process. The specific process is as follows:

[0073] Set the statistical sliding window W s and the detection sliding window W d , where W s Length L s , W d Length L d , I i-Ls-1 and I i-1 are the starting point and end point of the current statistical sliding window respectively, calculate the mean value μ and standard deviation σ of the current effective value of the statistical sliding window, and calculate the detection sliding window W d The change statistic Z-score on:

[0074]

[0075] Use the sliding window to traverse the current effective value sequence and find that Z is less than the threshold Z threshold The point values ​​are set to zero. After obtaining the Z value sequence, the sequence is divided into N categories through density clustering, where the corresponding point sequence in each category is a transient event process T.

[0076] For the transient event process T obtained, bilateral correction processing is performed by calculating the mean drift distance and slope. The point with the largest mean drift distance near the starting point of the transient event is corrected as the starting point, and the point where the slope is less than the slope threshold for the first time before the end point of the transient event is corrected as the end point, so as to obtain the accurate transient event starting point I start With stop point I end .

[0077] Starting Point I start With stop point I end The transient section of the event is T, starting point I start O before n The period constitutes the pre-event steady-state period O b , stop point I end After O n The period constitutes the post-event steady-state period O a .

[0078] In the embodiment of the present invention, the transient event process T is as follows: Figure 2 As shown, the smooth line segment represents the effective current value of the transient segment, and the segments with triangle lines before and after represent the effective current value of the steady-state segment before and after the event.

[0079] Step 3: Based on the target indicator sequence S of the event transient section, extract the path signature characteristics of the target indicator sequence S. If the collected load data includes high-frequency voltage data and high-frequency current data, go to step 4, otherwise go to step 5.

[0080] In order to obtain features suitable for deep learning model input representation, the path signature features of the target indicator sequence S are extracted. The path signature is the calculation result of the rough path theory and is currently widely used in the fields of machine learning and time series analysis.

[0081] For a scalar field F: Then for the path The path integral of is:

[0082]

[0083] where r:[a,b]→X is the parametric equation of the path.

[0084] For n-th path signature items, there are:

[0085]

[0086] Path X:[a,b]→R n The path signature S(X) a,b , is an infinitely long sequence consisting of signature terms of various orders:

[0087]

[0088] Since the path signature is an infinitely long sequence, it is not suitable for direct use in machine learning algorithms. Therefore, only the n-th order and previous signature items in the sequence are taken, and this new sequence is called the n-th order truncated path signature.

[0089] The active power P, reactive power Q or current effective value I of the transient section can be set to form the target indicator sequence S, and the n-order truncated path signature S of the sequence S can be extracted. T Used as features for machine learning input.

[0090] In the embodiment of the present invention, a two-dimensional target sequence S=[P, Q] is set, and a fourth-order truncated path signature is taken as a feature. The specific process is as follows:

[0091] Calculate S = [P, Q] size is (2, len)

[0092] Generate an equidistant sequence X of length len in the interval [0,1] path , forming a path X = [X path ,P,Q"

[0093] Calculate the 4th-order truncated path signature S(X) of path X, with a feature length of 3+3 2 +3 3 +3 4 =120, the extracted path signature feature is recorded as S T .

[0094] Step 4: Extract VI trajectory image features based on the high-frequency voltage and high-frequency current in the steady-state section before and after the event:

[0095] Extract the steady-state segment before the event O b The high-frequency voltage and current data are recorded as v b andi b .

[0096] Extract the post-event steady-state segment O a The high-frequency voltage and current data are recorded as v a andi a .

[0097] v b 、i b 、v a andi a Smoothing is performed to replace outliers with smoothed values, thereby avoiding the impact of outliers.

[0098] v b 、i b 、v a andi aPerform three Hermite interpolation fits respectively to obtain the interpolation function f vb 、f ib 、f va and f ia .

[0099] v b 、i b 、v a andi a Perform fast Fourier transform respectively to get the fundamental wave phase angle and And calculate the displacement of the waveform when the fundamental phase angle is zero, and get the initial sampling time t vb ,t ib ,t va and t ia , in the interpolation function f vb 、f ib 、f va 、f ia t vb ,t ib ,t va ,t ia Resample the same number of samples f at the initial sampling time, where f is O n To ensure that the sampling points of each cycle are aligned, the phase angle of v′ is aligned after resampling. b , i′ b , v′ a and i′ a .

[0100] v′ b , i′ b , v′ a and i′ a Medium n The average value of the sampling points is calculated for each cycle, and v is obtained. avr b 、i avr b 、v avr a 、i avr b .

[0101] Generate voltage sequence v ob and current sequence i ob :

[0102]

[0103] i ob =i avr a -i avr b

[0104] With voltage sequence v ob As the horizontal axis, the current sequence i ob Generate a VI trajectory graph for the vertical axis. The process is as follows:

[0105] Set the image pixel matrix A T Size (M, M), the initial pixel value is (255,255,255);

[0106] v ob 、i ob Resampled to M points, the pixel position of the jth voltage and current (x j ,y j )for:

[0107]

[0108]

[0109] Where int is the rounding function.

[0110] make

[0111]

[0112] If the jth voltage and current are in an upward trend in the voltage cycle, the pixel position (x j ,y j ) is taken as (value,value,255). If the jth voltage and current are in a downward trend in the voltage cycle, the pixel position (x j ,y j ) is taken as (value,value,value). The extracted image features are recorded as A T .

[0113] In the embodiment of the present invention, Figure 3 The figure shows the VI trajectory diagram for the microwave oven on event, where the gray trajectory represents the trajectory when the voltage cycle is rising, and the black trajectory represents the trajectory when the voltage cycle is falling.

[0114] Step 5: Select and extract the corresponding load imprint based on data conditions, and use the twin network framework for load identification.

[0115] If the collected load data includes high-frequency voltage and current data, the path signature feature S is selected. T and VI trajectory image feature A T As a load imprint, build a load imprint library of target device events.

[0116] If the collected load data does not include high-frequency voltage and current data, the path signature feature S is selected. TAs a load imprint, build a load imprint library of target device events.

[0117] If the collected load data includes high-frequency voltage and current data, the path signature feature S is selected. T and VI trajectory image feature A T As the load mark, denoted as F T =[S T , A T ].

[0118] If the collected load data does not include high-frequency voltage and current data, the path signature feature S is selected. T As the load mark, denoted as F T =S T .

[0119] Perform pairwise transformation on features. If T i With T j Belong to the same type of event, then the label y ij is 1, otherwise the label y ij is 0, and the sample form is as follows in For input 1, For input 2, y ij is the output.

[0120] In the embodiment of the present invention, the twin network structure is as follows: Figure 4 As shown, the twin neural network has two inputs. The two inputs are sent to two neural networks. The two neural networks with shared weights map the inputs to new spaces respectively, forming representation vectors of the inputs in the new space. By calculating the loss function, the similarity of the two inputs is evaluated as the output. Among them, if the collected load data does not include high-frequency voltage and current data, the shared weight neural network structure is constructed as follows Figure 5 As shown, the path signature feature S is transformed into T A representation vector is formed in the new space; if the collected load data includes high-frequency voltage and current data, a shared weight neural network structure is constructed such as Figure 6 As shown, by taking the path signature feature S T and VI trajectory image feature A T The mid-end fusion of multimodal data is performed through the fully connected layer and the convolutional layer to form a representation vector.

[0121] The trained twin network model is used for load identification. The process is as follows:

[0122] Takes as input the payload imprint of the detected unknown device event The load imprint library has K load imprints. The load imprints in the load imprint library are traversed and the load imprints in the load imprint library are used as input in sequence. Use the trained twin network model to get the output y ij , so there are K outputs y for the load imprint of the unknown device event ij , if the output y ij If the maximum value is greater than the threshold, the output y ij The maximum value corresponds to The event of the corresponding device is used as the load identification result, otherwise the event type is identified as an unknown event.

[0123] The present invention detects transient events by the change of the effective value of the current. It can accurately detect the starting and ending points of the total load transient events to distinguish the transient section and the steady-state section. Based on the steady-state section and the transient section before and after the electrical equipment event, the imprint features required for electrical equipment identification are accurately obtained, thereby completing load identification. Specifically, in a specific experimental scenario of the present invention, the power consumption data within a period of time is obtained by installing a smart meter on the laboratory open circuit bus. There are 5 types of loads, namely: refrigerator, water dispenser, kettle, microwave oven, printer.

[0124] The smart meter acquisition frequency is 16Hz, and the test time scale is 3 days. According to the method in step 2, a total of 615 events were detected, of which 609 were valid events. According to step 3, the path signature features were extracted, and the feature size was 120 dimensions. According to step 4, the VI trajectory features were extracted, and the feature size was (32, 32, 3). The trained twin network was used to calculate the similarity based on the load imprint library, and the load identification results were statistically obtained as follows:

[0125]

[0126]

[0127] Since there are fewer load categories, different loads will not run at the same time most of the time. When the load runs alone, the accuracy of using both features is relatively high. When different load events occur at the same time, the accuracy based on the path signature feature will drop significantly, while the accuracy based on the fusion feature will not be significantly affected.

[0128] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For researchers in the relevant field, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from them are still within the protection scope of the present invention.

Claims

1. A non-intrusive load identification method based on multimodal features, characterized in that: The following steps are involved: Step 1, load data acquisition; Step 2, performing event detection based on the power or current effective value, extracting the event transient section and the steady-state sections before and after the event transient section; Step 3: based on the target indicator sequence of the transient section of the event, extract the path signature features of the target indicator sequence. If the collected load data includes high-frequency voltage data and high-frequency current data, go to step 4; otherwise, go to step 5. Step 4, extracting VI trajectory image features based on the high-frequency voltage and high-frequency current in the steady-state section before and after the event; Step 5: Select and extract the corresponding load imprint based on the data conditions, and use the twin network framework to perform load identification; In step 5, if the collected load data includes high-frequency voltage data and high-frequency current data, the path signature feature S is selected. T and VI trajectory image feature A T As a load imprint, build a load imprint library of the target device event; if the collected load data does not include high-frequency voltage data and high-frequency current data, select the path signature feature S T As the load imprint, build a load imprint library of the target device event; based on the load imprint library, use the twin network to calculate the similarity between the load imprint of the detected event and the load imprint in the load imprint library. If the highest similarity is greater than the threshold, the event type is identified as the event type belonging to the load imprint with the highest similarity, otherwise the event type is identified as an unknown event.

2. The non-intrusive load identification method based on multimodal features according to claim 1 is characterized in that: The specific method for acquiring the load data in step 1 is: the load data assumes that the frequency is at least greater than 50 Hz. If the frequency is greater than 1 kHz, the load data includes high-frequency voltage data, high-frequency current data, active power P, reactive power Q, and current effective value I; otherwise, the load data includes active power P, reactive power Q, and current effective value I.

3. The non-intrusive load identification method based on multimodal features according to claim 1 is characterized in that: In step 2, the specific method of using the effective value of current to detect events and extracting the transient section of the event and the steady-state sections before and after the transient section of the event is: performing event detection based on the composite sliding window and bilateral correction method, extracting the start and end points of the transient process, and the starting point I start With stop point I end The transient section of the event is T, starting point I start O before n The period constitutes the pre-event steady-state period O b , stop point I end After O n The period constitutes the post-event steady-state period O a .

4. The non-intrusive load identification method based on multimodal features according to claim 1 is characterized in that: In step 3, active power P, reactive power Q or current effective value I are set to form a target indicator sequence S. In order to obtain features suitable for deep learning model input representation, the path signature features of the target indicator sequence S are extracted, and the n-order truncated path signature S of the target indicator sequence S is extracted. T Used as features for machine learning input.

5. The non-intrusive load identification method based on multimodal features according to claim 1 is characterized in that: In step 4, the pre-event steady-state segment O is extracted. b and post-event steady-state segment O a The high-frequency voltage and current data are processed by outlier smoothing, interpolation fitting, Fourier transform alignment, and resampling to calculate the periodic average voltage sequence v ob and period difference current sequence i ob , get the VI trajectory feature A T .

6. The non-intrusive load identification method based on multimodal features according to claim 3 is characterized in that: The specific method of step 2 is as follows: (1) Set the statistical sliding window W s and the detection sliding window W d , where W s Length L s , W d Length L d , I i-Ls-1 and I i-1 are the starting point and end point of the current statistical sliding window respectively, calculate the mean value μ and standard deviation σ of the current effective value of the statistical sliding window, and calculate the detection sliding window W d The change statistic Z-score on: (2) Use the sliding window to traverse the current effective value sequence and find the value Z less than the threshold Z threshold The point values ​​are set to zero. After obtaining the Z value sequence, the sequence is divided into N categories by density clustering, where the corresponding point sequence in each category is a transient event process T; (3) The transient event process T is corrected bilaterally by calculating the mean drift distance and slope. The point with the largest mean drift distance near the starting point of the transient event is corrected as the starting point, and the point where the slope is less than the slope threshold for the first time before the end point of the transient event is corrected as the end point. The accurate transient event starting point I is obtained. start With stop point I end ; Starting point I start With stop point I end The transient section of the event is T, starting point I start O before n The period constitutes the pre-event steady-state period O b , stop point I end After O n The period constitutes the post-event steady-state period O a .

7. The non-intrusive load identification method based on multimodal features according to claim 1 is characterized in that: The specific method of step 5 is as follows: (1) If the collected load data includes high-frequency voltage data and high-frequency current data, the path signature feature S is selected T and VI trajectory image feature A T As a load imprint, build a load imprint library of the target device event; if the collected load data does not include high-frequency voltage data and high-frequency current data, select the path signature feature S T As a load imprint, build a load imprint library of target device events; (2) If the collected load data includes high-frequency voltage data and high-frequency current data, the path signature feature S is selected T and VI trajectory image feature A T As the load mark, denoted as F T =[S T , A T ]; If the collected load data does not include high-frequency voltage data and high-frequency current data, the path signature feature S is selected T As the load mark, denoted as F T =S T ; (3) Perform pairwise transformation on features. If T i With T j Belong to the same type of event, then the label y ij is 1, otherwise the label y ij is 0, and the sample form is as follows in For input 1, For input 2, y ij is the output; The two inputs are sent to the twin neural network. The two neural networks with shared weights map the inputs to the new space respectively, forming the representation vector of the input in the new space. The similarity of the two inputs is evaluated as the output through the calculation of the loss function. If the collected load data does not include high-frequency voltage data and high-frequency current data, the path signature feature S is converted into T In the new space, a representation vector is formed; if the collected load data includes high-frequency voltage data and high-frequency current data, the path signature feature S T and VI trajectory image feature A T The mid-end fusion of multimodal data is performed through the fully connected layer and the convolutional layer to form a representation vector; (4) Use the trained twin network model to identify the load. The process is as follows: Takes as input the payload imprint of the detected unknown device event The load imprint library has K load imprints. The load imprints in the load imprint library are traversed and the load imprints in the load imprint library are used as input in sequence. Use the trained twin network model to get the output y ij , there are K outputs y for the load imprint of unknown device events ij , if the output y ij If the maximum value is greater than the threshold, the output y ij The maximum value corresponds to The event of the corresponding device is used as the load identification result, otherwise the event type is identified as an unknown event.

Images