CN114325081A - Non-invasive load identification method based on multi-modal characteristics - Google Patents
Non-invasive load identification method based on multi-modal characteristics Download PDFInfo
- Publication number
- CN114325081A CN114325081A CN202111638259.5A CN202111638259A CN114325081A CN 114325081 A CN114325081 A CN 114325081A CN 202111638259 A CN202111638259 A CN 202111638259A CN 114325081 A CN114325081 A CN 114325081A
- Authority
- CN
- China
- Prior art keywords
- load
- event
- data
- value
- frequency
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 60
- 230000001052 transient effect Effects 0.000 claims abstract description 60
- 238000001514 detection method Methods 0.000 claims abstract description 9
- 230000004927 fusion Effects 0.000 claims abstract description 6
- 230000008569 process Effects 0.000 claims description 20
- 238000005070 sampling Methods 0.000 claims description 15
- 230000006870 function Effects 0.000 claims description 12
- 238000013528 artificial neural network Methods 0.000 claims description 11
- 238000010586 diagram Methods 0.000 claims description 8
- 239000013598 vector Substances 0.000 claims description 8
- 238000012952 Resampling Methods 0.000 claims description 7
- 230000002159 abnormal effect Effects 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 7
- 230000002146 bilateral effect Effects 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 6
- 238000012937 correction Methods 0.000 claims description 6
- 238000010801 machine learning Methods 0.000 claims description 6
- 238000009499 grossing Methods 0.000 claims description 4
- 239000002131 composite material Substances 0.000 claims description 3
- 238000013136 deep learning model Methods 0.000 claims description 3
- 238000006073 displacement reaction Methods 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 230000000630 rising effect Effects 0.000 claims description 3
- 239000004576 sand Substances 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims 1
- 230000000737 periodic effect Effects 0.000 claims 1
- 238000012544 monitoring process Methods 0.000 description 6
- 230000008859 change Effects 0.000 description 3
- 238000010276 construction Methods 0.000 description 2
- 230000005611 electricity Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012300 Sequence Analysis Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 239000013604 expression vector Substances 0.000 description 1
- 238000012806 monitoring device Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Images
Abstract
The invention provides a non-invasive load identification method based on multi-modal characteristics, which comprises the following steps: 1) acquiring load data; 2) event detection is carried out based on the power or current effective value, and an event transient section and steady-state sections before and after the event transient section are extracted; 3) extracting path signature characteristics of the target index sequence based on the target index sequence of the event transient section, and if the acquired load data comprises high-frequency voltage data and high-frequency current data, turning to the step 4, otherwise, turning to the step 5; 4) extracting VI track image features based on the high-frequency voltage and the high-frequency current of the steady-state sections before and after the event; 5) and selecting and extracting corresponding load marks based on data conditions, and carrying out load identification by using a twin network framework. According to the method, the transient state path signature characteristics of the low-frequency data and the VI track characteristics of the high-frequency data are correspondingly extracted according to the acquired data frequency, and the load identification is realized through a middle-end fusion mode of multi-mode characteristics and a twin network.
Description
Technical Field
The invention relates to a non-invasive load monitoring method.
Technical Field
The load monitoring system is an important ring for resident demand response and safe electricity utilization, and the current load monitoring systems are roughly divided into two categories, namely invasive and non-invasive. The traditional intrusive load monitoring system is characterized in that a sensor is arranged at each load to monitor the running condition of each load, and the traditional intrusive load monitoring system is difficult to be widely applied due to the characteristics of higher hardware economic cost, low system reliability, small coverage range of users and the like. A Non-Intrusive Load Monitoring (NILM) system, in which a Monitoring device is installed at an electric power inlet, so that each electric Load inside a user can be analyzed, and information such as switching and energy consumption of each electric device can be provided.
The conversion process of the working state of the electric equipment is called a transient event, the transient characteristic is unique transient characteristic information such as current, voltage and power change information at the moment of occurrence of the event in the process of aiming at the transient event, and the steady characteristic is information such as voltage, current, power and harmonic wave before and after the occurrence of the event. The characteristic information that the state of the consumer changes with is called load signature.
The construction of the load imprint is the key of non-intrusive load identification accuracy, the existing load imprint construction method comprises a power change value based on low-frequency data, a transient power curve based on the low-frequency data, a VI track characteristic based on high-frequency steady-state data and the like, and the transient characteristic and the steady-state characteristic can only represent a part of the characteristics of the load characteristic. Common load identification models include a K-nearest neighbor algorithm, a hidden Markov algorithm, a convolutional neural network and the like. Load identification accuracy based on a hidden Markov algorithm is generally low, a supervision model based on low-frequency load imprint is difficult to distinguish for power similar loads, and an identification method based on a VI track and a neural network has high accuracy, but high-frequency data required by the method has high requirements on hardware facilities and more label data required by model training.
Disclosure of Invention
The invention provides a non-invasive load identification method based on multi-modal characteristics, and aims to solve at least one of the technical problems in the prior art.
In order to solve the technical problems, the invention adopts the following technical scheme:
a non-invasive load identification method based on multi-modal characteristics utilizes load data collected at an electric power inlet, and comprises the following steps:
step one, acquiring load data. Acquiring collected load data, wherein the assumed frequency of the load data is at least greater than 50Hz, and if the frequency is greater than 1kHz, the load data comprises high-frequency voltage data, high-frequency current data, active power P, reactive power Q and a current effective value I; otherwise, the load data comprises active power P, reactive power Q and current effective value I.
And secondly, carrying out event detection based on the power or current effective value, and extracting an event transient section and steady-state sections before and after the event transient section.
The method comprises the following steps of utilizing a current effective value, carrying out event detection based on a composite sliding window and bilateral correction method, and extracting a transient process starting point and a transient process stopping point, wherein the specific process comprises the following steps:
setting a statistical sliding window WsAnd detecting the sliding window WdWherein W issLength Ls，WdLength Ld，Ii-Ls-1And Ii-1Respectively as the starting point and the end point of the current statistical sliding window, calculating the current effective value mean value mu and the standard deviation sigma of the statistical sliding window, and calculating and detecting the sliding window WdChange statistic Z score above:
traversing the current effective value sequence by using a sliding window to enable Z to be smaller than a threshold value ZthresholdThe point value of the Z value sequence is set to be zero, and after the Z value sequence is obtained, the sequence is divided by density clusteringIs N categories, wherein the corresponding point sequence in each category is a transient event process T.
Performing bilateral correction processing on the obtained transient event process T in a mode of calculating a mean shift distance and a slope, correcting a point with the maximum mean shift distance near the starting point of the transient event as the starting point, and correcting a point with the slope smaller than a slope threshold value for the first time before the stopping point of the transient event as the stopping point to obtain an accurate starting point I of the transient eventstartAnd stop point Iend。
Starting point IstartAnd stop point IendConstituting an event transient section T, starting point IstartFront OnOne cycle constitutes a pre-event steady-state section ObStopping point IendAfter OnOne cycle constitutes the post-event steady-state section Oa。
And step three, extracting the path signature characteristics of the target index sequence S based on the target index sequence S of the event transient section, if the acquired load data comprises high-frequency voltage data and high-frequency current data, turning to step four, and if not, turning to step five.
In order to obtain features suitable for deep learning model input representation, path signature features of a target index sequence S are extracted, and the path signatures are calculation results of a rough path theory.
For a scalar field F:
then for the path
The path integral of (a) is:
wherein r [ a, b ] → X are parametric equations for the paths.
For the n-path signature entries are:
route X [ a, b]→RnPath signature S (X)a,bIs an infinite sequence, which is composed of signature items of various orders:
setting active power P, reactive power Q or current effective value I of transient section to form target index sequence S, and extracting n-order cut-off path signature S of sequence STAs a feature for machine learning input.
Step four, extracting VI track image characteristics based on the high-frequency voltage and the high-frequency current of the steady-state sections before and after the event:
extraction of Pre-event Steady State section ObHigh frequency voltage data and current data, denoted vbAnd ib。
Extracting post-event steady-state section OaHigh frequency voltage data and current data, denoted vaAnd ia。
For vb、ib、vaAnd iaAnd performing smoothing processing to replace the abnormal value with a smooth value so as to avoid the influence of the abnormal value.
For vb、ib、vaAnd iaRespectively carrying out three times of Hermite interpolation fitting to obtain an interpolation function fvb、fib、fvaAnd fia。
For vb、ib、vaAnd iaRespectively carrying out fast Fourier transform to obtain fundamental wave phase angles
And
and calculating the displacement of the waveform when the fundamental wave phase angle is zero to obtain the sampling initial time tvb、tib、tvaAnd tiaAt the interpolation function fvb、fib、fva、fiaAbove by t respectivelyvb、tib、tva、tiaResampling the same number of samples f for the initial sampling instant, where f is OnTo ensure that sampling points of each period are aligned to obtain v 'with aligned phase angles after resampling'b、i′b、v′aAnd i'a。
To v'b、i′b、v′aAnd i'aMiddle OnCalculating average value of sampling points of the sampling points in each period to obtain vavr b、iavr b、vavr a、iavr b。
Generating a voltage sequence vobAnd current sequence iob：
iob＝iavr a-iavr b
By a sequence of voltages vobAs abscissa, with current sequence iobGenerating a VI trajectory graph for the ordinate by:
setting a picture pixel matrix ATSize (M, M), value of initialization pixel point is (255 );
for vob、iobPixel position (x) of j th voltage and current, which are M points respectivelyj，yj) Comprises the following steps:
where int is the rounding function.
Order to
If the jth voltage current is in the rising trend of the voltage period, the pixel position (x)j，yj) Is taken to be (value, value,255), the pixel position (x) if the jth voltage current is in a downward trend in the voltage cyclej，yj) Is taken as (value, value, value). The extracted picture features are marked as AT。
And fifthly, selecting and extracting corresponding load marks based on data conditions, and carrying out load identification by using a twin network framework.
If the collected load data comprises high-frequency voltage and current data, selecting a path signature characteristic STAnd VI track image feature ATAnd building a load mark library of the target equipment event as the load mark.
If the collected load data does not include high-frequency voltage and current data, selecting a path signature characteristic STAnd building a load mark library of the target equipment event as the load mark.
If the collected load data comprises high-frequency voltage and current data, selecting a path signature characteristic STAnd VI track image feature ATAs a load mark, is noted as FT＝[ST，AT]。
If the collected load data does not include high-frequency voltage and current data, selecting a path signature characteristic STAs a load mark, is noted as FT＝ST。
Converting the features in pairs, if TiAnd TjBelong to the same type of event, then label yijIs 1, otherwise label yijIs 0, constitutes a sample form such as
Wherein
Is input into1，
To input 2, yijIs the output.
The twin neural network has two inputs, the two inputs are sent to the two neural networks, the two neural networks sharing the weight map the inputs to a new space respectively to form a representation vector input in the new space, and the similarity of the two inputs is evaluated as output through calculation of a loss function. If the collected load data does not include high-frequency voltage and current data, the path signature characteristic S is obtained through the full-connection layerTForming a representation vector in the new space; if the collected load data comprises high-frequency voltage and current data, the path is signed by a signature characteristic STAnd VI track image feature ATAnd performing middle-end fusion of the multi-mode data through the full connection layer and the convolution layer to form a representation vector.
And (3) carrying out load identification by using the trained twin network model, wherein the process is as follows:
load stamping of detected unknown device events as input
The load mark library has K load marks, the load marks in the load mark library are traversed, and the load marks in the load mark library are sequentially used as input
Obtaining output y by using the trained twin network modelijThus there are K outputs y for the load signature of an unknown device eventijIf y is outputijIf the maximum value is greater than the threshold value, y will be outputijMaximum value corresponds to
And taking the device event as a load identification result, otherwise, identifying the event type as an unknown event.
The implementation of the invention has the following beneficial effects:
the method can automatically select the corresponding load mark based on the data condition, the transient state path signature characteristic can be extracted based on low-frequency data, the path track characteristic of the transient state process can be effectively represented, the VI track characteristic is extracted based on high-frequency data, the steady state voltage and current track characteristic before and after an event can be effectively represented, and the transient state and steady state characteristic of the event can be effectively and comprehensively represented by the path signature and the VI track fusion characteristic. The twin network is a method for learning with few samples, and for each type of equipment, only a few samples are labeled, so that good effect can be obtained. The invention reduces the requirements on data conditions in two dimensions of acquisition frequency and sample labels, and is a practical and effective load identification method.
Drawings
Fig. 1 is an overall schematic diagram of a non-invasive load identification method based on multi-modal features according to the present invention.
FIG. 2 is a schematic diagram of transient and steady state segments according to an embodiment of the present invention.
Fig. 3 is a VI trace diagram of a microwave oven opening event according to an embodiment of the present invention.
FIG. 4 is a diagram of a twin network framework according to an embodiment of the present invention.
Fig. 5 is a diagram of a network structure with path signature characteristics as input according to an embodiment of the present invention.
FIG. 6 is a diagram of a network architecture with multimodal features as inputs in accordance with an embodiment of the present invention.
Detailed Description
The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
Fig. 1 shows an execution flow of a non-invasive load identification method based on multi-modal features, which includes the following steps:
acquiring load data;
secondly, event detection is carried out based on the power or current effective value, and an event transient section and steady-state sections before and after the event transient section are extracted;
step three, extracting path signature characteristics of the target index sequence based on the target index sequence of the event transient section, if the acquired load data comprises high-frequency voltage data and high-frequency current data, turning to step 4, and otherwise, turning to step 5;
extracting VI track image characteristics based on the high-frequency voltage and the high-frequency current of the steady-state sections before and after the event;
and fifthly, selecting and extracting corresponding load marks based on data conditions, and identifying the load by using a twin network framework.
The specific description of the above steps is as follows:
step one, acquiring load data. Acquiring collected load data, wherein the assumed frequency of the load data is at least greater than 50Hz, and if the frequency is greater than 1kHz, the load data comprises high-frequency voltage data, high-frequency current data, active power P, reactive power Q and a current effective value I; otherwise, the load data comprises active power P, reactive power Q and current effective value I.
And secondly, carrying out event detection based on the power or current effective value, and extracting an event transient section and steady-state sections before and after the event transient section.
In the embodiment of the invention, event detection is carried out by using a current effective value based on a composite sliding window and bilateral correction method, and the starting point and the ending point of a transient process are extracted, wherein the specific process comprises the following steps:
setting a statistical sliding window WsAnd detecting the sliding window WdWherein W issLength Ls，WdLength Ld，Ii-Ls-1And Ii-1Respectively as the starting point and the end point of the current statistical sliding window, calculating the current effective value mean value mu and the standard deviation sigma of the statistical sliding window, and calculating and detecting the sliding window WdChange statistic Z score above:
traversing the current effective value sequence by using a sliding window to enable Z to be smaller than a threshold value ZthresholdThe point value of the time sequence is set to be zero, after the Z value sequence is obtained, the sequence is divided into N categories through density clustering, wherein the corresponding point sequence in each category is a transient event process T.
Performing bilateral correction processing on the obtained transient event process T in a mode of calculating a mean shift distance and a slope, correcting a point with the maximum mean shift distance near the starting point of the transient event as the starting point, and correcting a point with the slope smaller than a slope threshold value for the first time before the stopping point of the transient event as the stopping point to obtain an accurate starting point I of the transient eventstartAnd stop point Iend。
Starting point IstartAnd stop point IendConstituting an event transient section T, starting point IstartFront OnOne cycle constitutes a pre-event steady-state section ObStopping point IendAfter OnOne cycle constitutes the post-event steady-state section Oa。
In the embodiment of the present invention, the transient event process T is shown in fig. 2, wherein the smooth line segments represent the current effective values of the transient segments, and the front and back triangular line segments represent the current effective values of the steady-state segments before and after the event.
And step three, extracting the path signature characteristics of the target index sequence S based on the target index sequence S of the event transient section, if the acquired load data comprises high-frequency voltage data and high-frequency current data, turning to step four, and if not, turning to step five.
In order to obtain features suitable for deep learning model input representation, the path signature features of a target index sequence S are extracted, and the path signature is a calculation result of a rough path theory and is widely applied to the field of machine learning and time sequence analysis at present.
For a scalar field F:
then for the path
The path integral of (a) is:
wherein r [ a, b ] → X are parametric equations for the paths.
For the n-path signature entries are:
route X [ a, b]→RnPath signature S (X)a,bIs an infinite sequence, which is composed of signature items of various orders:
since the path signature is an infinitely long sequence, it is not suitable for direct use in machine learning algorithms. Thus, only the n-th order and previous signature entries in the sequence are taken, and the new sequence is called an n-th order truncated path signature.
The active power P, the reactive power Q or the current effective value I of the transient section can be set to form a target index sequence S, and an n-order cut-off path signature S of the sequence S is extractedTAs a feature for machine learning input.
In the embodiment of the present invention, a two-dimensional target sequence S ═ P, Q is set, and a 4-order truncated path signature is taken as a feature, and the specific process is as follows:
calculate S ═ P, Q ] size (2, len)
In the interval [0,1]Generation of equidistant sequence X of len lengthpathForm a path X ═ Xpath,P,Q」
Calculating 4-order truncation path signature S (X) characteristic of the path X, wherein the characteristic length is 3+32+33+34120, the extracted path signature feature is denoted as ST。
Step four, extracting VI track image characteristics based on the high-frequency voltage and the high-frequency current of the steady-state sections before and after the event:
extraction of Pre-event Steady State section ObHigh frequency voltage data and current data, denoted vbAnd ib。
Extracting post-event steady-state section OaHigh frequency voltage data and current data, denoted vaAnd ia。
For vb、ib、vaAnd iaAnd performing smoothing processing to replace the abnormal value with a smooth value so as to avoid the influence of the abnormal value.
For vb、ib、vaAnd iaRespectively carrying out three times of Hermite interpolation fitting to obtain an interpolation function fvb、fib、fvaAnd fia。
For vb、ib、vaAnd iaRespectively carrying out fast Fourier transform to obtain fundamental wave phase angles
And
and calculating the displacement of the waveform when the fundamental wave phase angle is zero to obtain the sampling initial time tvb、tib、tvaAnd tiaAt the interpolation function fvb、fib、fva、fiaAbove by t respectivelyvb、tib、tva、tiaResampling the same number of samples f for the initial sampling instant, where f is OnTo ensure that sampling points of each period are aligned to obtain v 'with aligned phase angles after resampling'b、i′b、v′aAnd i'a。
To v'b、i′b、v′aAnd i'aMiddle OnCalculating average value of sampling points of the sampling points in each period to obtain vavr b、iavr b、vavr a、iavr b。
Generating a voltage sequence vobAnd current sequence iob：
iob＝iavr a-iavrb
By a sequence of voltages vobAs abscissa, with current sequence iobGenerating a VI trajectory graph for the ordinate by:
setting a picture pixel matrix ATSize (M, M), value of initialization pixel point is (255 );
for vob、iobPixel position (x) of j th voltage and current, which are M points respectivelyj，yj) Comprises the following steps:
where int is the rounding function.
Order to
If the jth voltage current is in the rising trend of the voltage period, the pixel position (x)j，yj) Is taken to be (value, value,255), the pixel position (x) if the jth voltage current is in a downward trend in the voltage cyclej，yj) Is taken as (value, value, value). The extracted picture features are marked as AT。
In an embodiment of the present invention, fig. 3 is a VI trace diagram of a microwave oven starting event, wherein a gray trace represents a trace when a voltage period rises, and a black trace represents a trace when a voltage period falls.
And fifthly, selecting and extracting corresponding load marks based on data conditions, and carrying out load identification by using a twin network framework.
If the collected load data comprises high-frequency voltage and current data, selecting a path signature characteristic STAnd VI track image feature ATAnd building a load mark library of the target equipment event as the load mark.
If the collected load data does not include high-frequency voltage and current data, selecting a path signature characteristic STAnd building a load mark library of the target equipment event as the load mark.
If the collected load data comprises high-frequency voltage and current data, selecting a path signature characteristic STAnd VI track image feature ATAs a load mark, is noted as FT＝[ST，AT]。
If the collected load data does not include high-frequency voltage and current data, selecting a path signature characteristic STAs a load mark, is noted as FT＝ST。
Converting the features in pairs, if TiAnd TjBelong to the same type of event, then label yijIs 1, otherwise label yijIs 0, constitutes a sample form such as
Wherein
In order to input the value 1, the input value is input,
to input 2, yijIs the output.
In the embodiment of the present invention, the twin network structure is as shown in fig. 4, the twin neural network has two inputs, the two inputs are sent to the two neural networks, and the two neural networks sharing the weight map the inputs to new spaces respectively, so as to form the expression vector of the inputs in the new spaces. Through calculation of the loss function, the similarity of the two inputs is evaluated as an output. If the collected load data does not include high-frequency voltage and current data, a shared weight neural network structure is built as shown in fig. 5, and path signature characteristics S are obtained through a full connection layerTForming a representation vector in the new space; if the collected load data comprises high-frequency voltage and current data, constructing a shared weight neural network structure as shown in the figure6, by signing the path with a signature STAnd VI track image feature ATAnd performing middle-end fusion of the multi-mode data through the full connection layer and the convolution layer to form a representation vector.
And (3) carrying out load identification by using the trained twin network model, wherein the process is as follows:
load stamping of detected unknown device events as input
The load mark library has K load marks, the load marks in the load mark library are traversed, and the load marks in the load mark library are sequentially used as input
Obtaining output y by using the trained twin network modelijThus there are K outputs y for the load signature of an unknown device eventijIf y is outputijIf the maximum value is greater than the threshold value, y will be outputijMaximum value corresponds to
And taking the device event as a load identification result, otherwise, identifying the event type as an unknown event.
The transient event is detected according to the change of the current effective value, the starting point and the ending point of the total load transient event can be accurately detected to distinguish the transient section from the steady section, and the mark characteristics required by the electric equipment identification are accurately acquired based on the steady section and the transient section before and after the electric equipment event occurs, so that the load identification is completed. Specifically, in a specific experimental scene of the invention, the intelligent electric meter arranged on the laboratory air-on bus is used for obtaining electricity consumption data in a period of time, and the load is 5 types in total, namely a refrigerator, a water dispenser, a kettle, a microwave oven and a printer.
And the intelligent electric meter acquires the data with the frequency of 16Hz and the test time scale of 3 days, and 615 events are detected in total according to the method in the step two, wherein 609 effective events are detected. And extracting the path signature features according to the step three, wherein the feature size is 120 dimensions. And extracting VI track features according to the fourth step, wherein the feature size is (32, 32, 3). The similarity is calculated by using the trained twin network based on the load imprint library, and the obtained load identification result statistics are as follows:
because the load types are less, different loads can not run simultaneously in most of time, the accuracy rate of using two characteristics is higher when the loads run independently, when different loads happen simultaneously, the accuracy rate based on the path signature characteristics can be obviously reduced, and the accuracy rate based on the fusion characteristics is not obviously influenced.
It should be understood that the above examples are only for clarity of illustration and are not intended to limit the embodiments. Variations or modifications in other variations may occur to those skilled in the art based upon the foregoing description. This need not be, nor should all embodiments be exhaustive. And obvious variations or modifications of the invention may be made without departing from the scope of the invention.
Claims (10)
1. A non-invasive load identification method based on multi-modal characteristics is characterized by comprising the following steps:
step 1, acquiring load data;
step 2, event detection is carried out based on the power or current effective value, and an event transient section and steady-state sections before and after the event transient section are extracted;
step 3, based on the target index sequence of the event transient section, extracting the path signature characteristics of the target index sequence, if the acquired load data comprises high-frequency voltage data and high-frequency current data, turning to step 4, and otherwise, turning to step 5;
step 4, extracting VI track image characteristics based on high-frequency voltage and high-frequency current of steady-state sections before and after an event;
and 5, selecting and extracting corresponding load marks based on data conditions, and carrying out load identification by using a twin network framework.
2. The method of claim 1, wherein the method comprises the following steps: the specific method for acquiring the load data in the step 1 comprises the following steps: the load data is supposed to have a frequency at least greater than 50Hz, and if the frequency is greater than 1kHz, the load data comprises high-frequency voltage data, high-frequency current data, active power P, reactive power Q and a current effective value I; otherwise, the load data comprises active power P, reactive power Q and current effective value I.
3. The method of claim 1, wherein the method comprises the following steps: in the step 2, event detection is performed by using the current effective value, and a specific method for extracting the event transient section and the steady-state sections before and after the event transient section is as follows: event detection is carried out based on a composite sliding window and bilateral correction method, and a starting point, a stopping point and a starting point I of a transient process are extractedstartAnd stop point IendConstituting an event transient section T, starting point IstartFront OnOne cycle constitutes a pre-event steady-state section ObStopping point IendAfter OnOne cycle constitutes the post-event steady-state section Oa。
4. The method of claim 1, wherein the method comprises the following steps: in the step 3, the active power P, the reactive power Q or the current effective value I are set to form a target index sequence S, in order to obtain the characteristics suitable for the input representation of the deep learning model, the path signature characteristics of the target index sequence S are extracted, and the n-order cut-off path signature S of the target index sequence S is extractedTAs a feature for machine learning input.
5. The method of claim 1 based on multi-modal featuresThe non-intrusive load identification method is characterized in that: in the step 4, the steady-state section O before the event is extractedbAnd post event steady state region OaThe periodic average voltage sequence v is calculated by carrying out abnormal value smoothing processing, interpolation fitting, Fourier transform phase angle alignment and resampling on the high-frequency voltage data and the current dataobAnd period difference current sequence iobObtaining VI locus diagram characteristic AT。
6. The method of claim 1, wherein the method comprises the following steps: in the step 5, if the collected load data includes high-frequency voltage data and high-frequency current data, the path signature feature S is selectedTAnd VI track image feature ATAs a load mark, building a load mark library of a target device event; if the collected load data does not include high-frequency voltage data and high-frequency current data, selecting a path signature characteristic STAs a load mark, building a load mark library of a target device event; and based on the load stamp library, performing similarity calculation on the load stamps of the detected events and the load stamps in the load stamp library by using a twin network, if the highest similarity is greater than a threshold value, identifying the event type as the event type to which the highest-similarity load stamp belongs, and otherwise, identifying the event type as an unknown event.
7. The method for non-invasive load identification based on multi-modal features according to claim 3, wherein the specific method of the step 2 is as follows:
(1) setting a statistical sliding window WsAnd detecting the sliding window WdWherein W issLength Ls，WdLength Ld，Ii-Ls-1And Ii-1Respectively as the starting point and the end point of the current statistical sliding window, calculating the current effective value mean value mu and the standard deviation sigma of the statistical sliding window, and calculating and detecting the sliding window WdChange statistic Z score above:
(2) traversing the current effective value sequence by using a sliding window to enable Z to be smaller than a threshold value ZthresholdSetting the point value of the point sequence to be zero, dividing the sequence into N categories through density clustering after obtaining a Z value sequence, wherein the corresponding point sequence in each category is a transient event process T;
(3) performing bilateral correction processing on the obtained transient event process T in a mode of calculating a mean shift distance and a slope, correcting a point with the maximum mean shift distance near the starting point of the transient event as the starting point, and correcting a point with the slope smaller than a slope threshold value for the first time before the stopping point of the transient event as the stopping point to obtain an accurate starting point I of the transient eventstartAnd stop point Iend(ii) a Starting point IstartAnd stop point IendConstituting an event transient section T, starting point IstartFront OnOne cycle constitutes a pre-event steady-state section ObStopping point IendAfter OnOne cycle constitutes the post-event steady-state section Oa。
8. The method for non-invasive load identification based on multi-modal features according to claim 4, wherein the specific method of the step 3 is as follows:
(1) for scalar fields
Then for the path
The path integral of (a) is:
wherein r [ a, b ] → X are parametric equations of the paths;
(2) for the n-path signature entries are:
route X [ a, b]→RnPath signature S (X)a,bIs an infinite sequence, which is composed of signature items of various orders:
(3) setting active power P, reactive power Q or current effective value I of transient section to form target index sequence S, and extracting n-order cut-off path signature S of sequence STAs a feature for machine learning input.
9. The method for non-invasive load identification based on multi-modal features as claimed in claim 5, wherein the specific method of step 4 is as follows:
(1) extraction of Pre-event Steady State section ObHigh-frequency voltage data and high-frequency current data, denoted as vbAnd ib(ii) a Extracting post-event steady-state section OaHigh-frequency voltage data and high-frequency current data, denoted as vaAnd ia；
(2) For vb、ib、vaAnd iaCarrying out smoothing processing, and replacing the abnormal value with a smooth value so as to avoid the influence of the abnormal value;
(3) for vb、ib、vaAnd iaRespectively carrying out three times of Hermite interpolation fitting to obtain an interpolation function fvb、fib、fvaAnd fia；
(4) For vb、ib、vaAnd iaRespectively carrying out Fourier transform to obtain fundamental wave phase angles
And
and calculating the displacement of the waveform when the fundamental wave phase angle is zero to obtain the sampling initial time tvb、tib、tvaAnd tiaAt the interpolation function fvb、fib、fva、fiaAbove by t respectivelyvb、tib、tva、tiaResampling the same number of samples f for the initial sampling instant, where f is OnTo ensure that sampling points of each period are aligned to obtain v 'with aligned phase angles after resampling'b、i′b、v′aAnd i'a；
(5) To v'b、i′b、v′aAnd i'aMiddle OnCalculating average value of sampling points of the sampling points in each period to obtain vavr b、iavr b、vavr a、iavr b；
(6) Generating a voltage sequence vobAnd current sequence iob：
iob＝iavr a-iavr b
(7) By a sequence of voltages vobAs abscissa, with current sequence iobGenerating a VI trajectory graph for the ordinate by:
setting a picture pixel matrix ATSize (M, M), value of initialization pixel point is (255 );
(8) for vob、iobPixel position (x) of j th voltage and current, which are M points respectivelyj，yj) Comprises the following steps:
wherein int is a rounding function;
order to
If the jth voltage current is in the rising trend of the voltage period, the pixel position (x)j，yj) Is taken to be (value, value,255), the pixel position (x) if the jth voltage current is in a downward trend in the voltage cyclej，yj) Takes the value of (value, value, value); the extracted picture features are marked as AT。
10. The method for non-invasive load identification based on multi-modal features according to claim 6, wherein the specific method of the step 5 is as follows:
(1) if the collected load data comprises high-frequency voltage data and high-frequency current data, selecting a path signature characteristic STAnd VI track image feature ATAs a load mark, building a load mark library of a target device event; if the collected load data does not include high-frequency voltage data and high-frequency current data, selecting a path signature characteristic STAs a load mark, building a load mark library of a target device event;
(2) if the collected load data comprises high-frequency voltage data and high-frequency current data, selecting a path signature characteristic STAnd VI track image feature ATAs a load mark, is noted as FT＝[ST，AT](ii) a If the collected load data does not include high-frequency voltage data and high-frequency current data, selecting a path signature characteristic STAs a load mark, is noted as FT＝ST；
(3) Converting the features in pairs, if TiAnd TjIf the events belong to the same type, then the mark is markedSign yijIs 1, otherwise label yijIs 0, constitutes a sample form such as
Wherein
In order to input the value 1, the input value is input,
to input 2, yijIs an output;
sending the two inputs into a twin neural network, respectively mapping the inputs into a new space by the two weight sharing neural networks to form a representation vector input into the new space, and evaluating the similarity of the two inputs as output through calculation of a loss function; if the collected load data does not include high-frequency voltage data and high-frequency current data, the path signature characteristic S is signed through the full connection layerTForming a representation vector in the new space; if the collected load data comprises high-frequency voltage data and high-frequency current data, the path is signed by a signature characteristic STAnd VI track image feature ATPerforming middle-end fusion of multi-mode data through the full-connection layer and the convolution layer to form a representation vector;
(4) and (3) carrying out load identification by using the trained twin network model, wherein the process is as follows:
load stamping of detected unknown device events as input
The load mark library has K load marks, the load marks in the load mark library are traversed, and the load marks in the load mark library are sequentially used as input
Obtaining output y by using the trained twin network modelijThere are K outputs y for the load signature of an unknown device eventijIf y is outputijMaximum value is largeAt threshold, then y will be outputijMaximum value corresponds to
And taking the device event as a load identification result, otherwise, identifying the event type as an unknown event.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111638259.5A CN114325081A (en) | 2021-12-29 | 2021-12-29 | Non-invasive load identification method based on multi-modal characteristics |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111638259.5A CN114325081A (en) | 2021-12-29 | 2021-12-29 | Non-invasive load identification method based on multi-modal characteristics |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114325081A true CN114325081A (en) | 2022-04-12 |
Family
ID=81016389
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111638259.5A Pending CN114325081A (en) | 2021-12-29 | 2021-12-29 | Non-invasive load identification method based on multi-modal characteristics |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114325081A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116595488A (en) * | 2023-07-19 | 2023-08-15 | 青岛鼎信通讯股份有限公司 | Non-invasive load identification method based on intelligent ammeter |
Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2014075581A1 (en) * | 2012-11-16 | 2014-05-22 | 天津大学 | Current-mode matching method for non-invasive power load monitoring and decomposition |
| US20190140445A1 (en) * | 2016-05-18 | 2019-05-09 | Mitsubishi Electric Corporation | Current/voltage control system |
| CN109813978A (en) * | 2018-12-25 | 2019-05-28 | 武汉中原电子信息有限公司 | A kind of non-intruding load-type recognition methods of variation characteristic between comprehensive transient characteristic and stable state |
| CN110514889A (en) * | 2019-07-19 | 2019-11-29 | 浙江万胜智能科技股份有限公司 | A kind of method and system of non-intrusion type household electricity remained capacity |
| CN110956220A (en) * | 2019-12-11 | 2020-04-03 | 深圳市活力天汇科技股份有限公司 | Non-invasive household appliance load identification method |
| CN111027408A (en) * | 2019-11-19 | 2020-04-17 | 广西电网有限责任公司电力科学研究院 | Load identification method based on support vector machine and V-I curve characteristics |
| CN111898637A (en) * | 2020-06-28 | 2020-11-06 | 南京工程学院 | Feature selection algorithm based on Relieff-DDC |
| CN112418722A (en) * | 2020-12-08 | 2021-02-26 | 浙江大学 | Non-invasive load identification method based on V-I (velocity-amplitude) trajectory graph and neural network |
| CN112821559A (en) * | 2021-01-22 | 2021-05-18 | 西安理工大学 | Non-invasive household appliance load depth re-identification method |
| CN113255236A (en) * | 2021-07-07 | 2021-08-13 | 浙江大学 | Non-invasive load self-adaptive identification method based on twin network |
| CN113505241A (en) * | 2021-07-15 | 2021-10-15 | 润建股份有限公司 | Intelligent diagnosis method for potential safety hazards of electricity utilization based on knowledge graph |
| CN113569445A (en) * | 2021-07-01 | 2021-10-29 | 洛阳星派数值仿真研究院有限公司 | Steel structure health monitoring system and method based on digital twinning technology |
-
2021
- 2021-12-29 CN CN202111638259.5A patent/CN114325081A/en active Pending
Patent Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2014075581A1 (en) * | 2012-11-16 | 2014-05-22 | 天津大学 | Current-mode matching method for non-invasive power load monitoring and decomposition |
| US20190140445A1 (en) * | 2016-05-18 | 2019-05-09 | Mitsubishi Electric Corporation | Current/voltage control system |
| CN109813978A (en) * | 2018-12-25 | 2019-05-28 | 武汉中原电子信息有限公司 | A kind of non-intruding load-type recognition methods of variation characteristic between comprehensive transient characteristic and stable state |
| CN110514889A (en) * | 2019-07-19 | 2019-11-29 | 浙江万胜智能科技股份有限公司 | A kind of method and system of non-intrusion type household electricity remained capacity |
| CN111027408A (en) * | 2019-11-19 | 2020-04-17 | 广西电网有限责任公司电力科学研究院 | Load identification method based on support vector machine and V-I curve characteristics |
| CN110956220A (en) * | 2019-12-11 | 2020-04-03 | 深圳市活力天汇科技股份有限公司 | Non-invasive household appliance load identification method |
| CN111898637A (en) * | 2020-06-28 | 2020-11-06 | 南京工程学院 | Feature selection algorithm based on Relieff-DDC |
| CN112418722A (en) * | 2020-12-08 | 2021-02-26 | 浙江大学 | Non-invasive load identification method based on V-I (velocity-amplitude) trajectory graph and neural network |
| CN112821559A (en) * | 2021-01-22 | 2021-05-18 | 西安理工大学 | Non-invasive household appliance load depth re-identification method |
| CN113569445A (en) * | 2021-07-01 | 2021-10-29 | 洛阳星派数值仿真研究院有限公司 | Steel structure health monitoring system and method based on digital twinning technology |
| CN113255236A (en) * | 2021-07-07 | 2021-08-13 | 浙江大学 | Non-invasive load self-adaptive identification method based on twin network |
| CN113505241A (en) * | 2021-07-15 | 2021-10-15 | 润建股份有限公司 | Intelligent diagnosis method for potential safety hazards of electricity utilization based on knowledge graph |
Non-Patent Citations (2)
| Title |
|---|
| NILS HUXOLL 等: "Model Identification and Parameter Tuning of Dynamic Loads in Power Distribution Grid: Digital Twin Approach", 《2021 INTERNATIONAL CONFERENCE ON SMART ENERGY SYSTEMS AND TECHNOLOGIES》, 27 September 2021 (2021-09-27), pages 1 - 6, XP033977771, DOI: 10.1109/SEST50973.2021.9543095 * |
| 雷怡琴 等: "电力系统负荷非侵入式监测方法研究", 《电工技术学报》, vol. 36, no. 11, 30 June 2021 (2021-06-30), pages 2288 - 2297 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116595488A (en) * | 2023-07-19 | 2023-08-15 | 青岛鼎信通讯股份有限公司 | Non-invasive load identification method based on intelligent ammeter |
| CN116595488B (en) * | 2023-07-19 | 2023-11-14 | 青岛鼎信通讯股份有限公司 | Non-invasive load identification method based on intelligent ammeter |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | Non-intrusive load monitoring algorithm based on features of V–I trajectory | |
| Chen et al. | Intelligent home-appliance recognition over IoT cloud network | |
| Shi et al. | Nonintrusive load monitoring in residential households with low-resolution data | |
| CN110400231B (en) | Failure rate estimation method for electric energy metering equipment based on weighted nonlinear Bayes | |
| CN111027408A (en) | Load identification method based on support vector machine and V-I curve characteristics | |
| CN111628494B (en) | Low-voltage distribution network topology identification method and system based on logistic regression method | |
| Li et al. | A nonintrusive load identification model based on time-frequency features fusion | |
| Iksan et al. | Appliances identification method of non-intrusive load monitoring based on load signature of VI trajectory | |
| CN111009893B (en) | Household user short-term load prediction method based on load decomposition technology | |
| CN110569876A (en) | Non-invasive load identification method and device and computing equipment | |
| CN112598303A (en) | Non-invasive load decomposition method based on combination of 1D convolutional neural network and LSTM | |
| Wu et al. | A load identification algorithm of frequency domain filtering under current underdetermined separation | |
| CN112435142A (en) | Power load identification method and load power utilization facility knowledge base construction method thereof | |
| CN110031680A (en) | A kind of system side harmonic impedance estimation method and system | |
| CN111932051A (en) | Malicious behavior detection method based on non-invasive power terminal time sequence monitoring | |
| CN114325081A (en) | Non-invasive load identification method based on multi-modal characteristics | |
| CN116127409A (en) | Non-invasive load identification method based on Gram angle difference field feature fusion | |
| Schirmer et al. | Double Fourier integral analysis based convolutional neural network regression for high-frequency energy disaggregation | |
| Qu et al. | A residual convolutional neural network with multi-block for appliance recognition in non-intrusive load identification | |
| CN109993424B (en) | Non-interference type load decomposition method based on width learning algorithm | |
| Jing et al. | Adjustable piecewise regression strategy based wind turbine power forecasting for probabilistic condition monitoring | |
| CN105259409B (en) | A kind of frequency power signal computational methods based on zero passage time domain specification | |
| CN104182910A (en) | Correlation-associated wind power output scene construction method | |
| CN116738866A (en) | Instant learning soft measurement modeling method based on time sequence feature extraction | |
| Zhang et al. | An autocorrelation incremental fuzzy clustering framework based on dynamic conditional scoring model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |