CN109813978A - A kind of non-intruding load-type recognition methods of variation characteristic between comprehensive transient characteristic and stable state - Google Patents

A kind of non-intruding load-type recognition methods of variation characteristic between comprehensive transient characteristic and stable state Download PDF

Info

Publication number
CN109813978A
CN109813978A CN201811594124.1A CN201811594124A CN109813978A CN 109813978 A CN109813978 A CN 109813978A CN 201811594124 A CN201811594124 A CN 201811594124A CN 109813978 A CN109813978 A CN 109813978A
Authority
CN
China
Prior art keywords
state
electrical appliance
transient
characteristic
stable state
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.)
Granted
Application number
CN201811594124.1A
Other languages
Chinese (zh)
Other versions
CN109813978B (en
Inventor
胡彦杰
周任飞
杨旸
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan Zhongyuan Electronic Information Co Ltd
Original Assignee
Wuhan Zhongyuan Electronic Information Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Wuhan Zhongyuan Electronic Information Co Ltd filed Critical Wuhan Zhongyuan Electronic Information Co Ltd
Priority to CN201811594124.1A priority Critical patent/CN109813978B/en
Publication of CN109813978A publication Critical patent/CN109813978A/en
Application granted granted Critical
Publication of CN109813978B publication Critical patent/CN109813978B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Abstract

The invention proposes a kind of non-intruding load-type recognition methods of the variation characteristic between comprehensive transient characteristic and stable state, in order to prevent since the close caused electrical appliance type of steady state characteristic and state distinguish problem, it realizes the effect of system intelligent Statistical error electrical appliance and improves the recognition effect of electrical appliance transient state and steady-state process, the present invention takes monitoring storage mode, after fundamental current amplitude changes, start to store the sampling period, after the sampling period of multiple variations is not detected, transient process identification and the steady-state process of electrical appliance are identified by improved CUSUM method again.Meanwhile the present invention pass through convolutional neural networks identify electrical appliance transient process and C means Method identify electrical appliance stable state between change procedure combine means, identify electrical appliance type and its which kind of working condition be in.In addition, the present invention can be the dynamic personalized electrical appliance property data base of user's building using the essential attribute of C means clustering algorithm.Finally, the knots modification of current amplitude under each frequency between more accurate front and back stable state can be obtained the present invention is based on sinusoidal Superposition Formula.

Description

A kind of non-intruding load-type identification of variation characteristic between comprehensive transient characteristic and stable state Method
Technical field
The present invention relates to a kind of the non-of the variation characteristic between smart grid field more particularly to comprehensive transient characteristic and stable state Invade load-type recognition methods.
Background technique
With the increase of world's extreme climate caused by the reduction of non-renewable energy resources and greenhouse effects, how to improve The service efficiency of the energy becomes the important ring in the policy of various countries' realization environmental protection and energy saving.Electric energy is as a kind of close with resident The relevant energy.In the past, resident can only understand total electricity consumption situation of family, it is difficult to understand by traditional ammeter The power consumption condition of each specific electrical appliance to targetedly select more energy-efficient electrical appliance, and then reduces Energy consumption.Therefore how to realize and allow user it can be seen that each electrical appliance consumes power levels, together as battery of mobile phone management When further utilize big data be resident recommend energy conservation and environmental protection, cheap and good-quality electrical appliance, be realize Intelligent energy-saving ring The key of guarantor.And with the development of computer technology, the research institution and company of every country have all started grinding for related fields Study carefully.
Currently, identifying that the mode of electrical appliance mainly takes intrusive mood and two kinds of non-intrusion type in the world.Although intrusive mood is known Other mode has the accurate advantage of recognition effect, but since the inconvenient and product price of device installation is high, causes this Mode is difficult to carry out extensive marketing.And non-intrusion type load monitoring (Non-Intrusive Load Monitoring, It is abbreviated as NILM) the installation monitoring equipment in the bus of residential electricity consumption is only needed, based on more perfect artificial of current development Intelligent recognition algorithm, cloud computing and big data distributed management technology can only pass through the electric current and electricity in home circuit bus The variation of pressure, maximum probability accurately identify electrical appliance type and state, effectively can provide detailed household electricity for user Each electrical appliance working condition in system receives the favor of most of research institution and company both at home and abroad.
Currently, it is to identify first to the stable state and transient state of electrical appliance work that most of NILM, which mainly takes, extracts use The steady state characteristic of electric appliance is analyzed.By acquiring a large amount of electrical appliance steady state characteristic, such as fundamental current, harmonic current, have Function power and reactive power establish the database of corresponding electrical appliance, later using the property data base of electrical appliance, using poly- The intelligent algorithms such as class, Hidden Markov Chain, neural network, support vector machine and particle swarm algorithm, to more in single steady-state process The state vector that a electrical appliance is likely to be at seeks optimal solution.
But above method has the following problems:
1. certain electrical appliances have extremely similar steady state characteristic in the steady state, the kind of electrical appliance can not be effectively distinguished Class;
2. the interdependency of pair acquisition database is higher, electrical appliance identification can not be optimized according to the variation of household power utilization environment Effect;
3. past research method is not due under ideal research environment, considering that electrical appliance state changes what event occurred Inner link between opportunity and sampled data, will lead to due to certain electrical appliances itself transient state have longer time, It is only capable of adopting in one sampling period and leads to provide the sampling sample of mistake to half of transient event for identifying system, lead to system Identify mistake.
Summary of the invention
In view of this, the invention proposes the changes between a kind of synthesis transient characteristic that electrical appliance recognition accuracy is high and stable state Change the non-intruding load-type recognition methods of feature.
The technical scheme of the present invention is realized as follows: the present invention provides the changes between a kind of comprehensive transient characteristic and stable state The non-intruding load-type recognition methods for changing feature, includes the following steps,
S1 takes monitoring storage mode, after fundamental current amplitude changes, starts to store the sampling period;
S2 after the sampling period of multiple variations is not detected, then identifies the transient process and steady-state process of electrical appliance, needle Transient process to the electrical appliance identified executes step S3;For the steady-state process of the electrical appliance identified, step is executed S4;
S3 is identified the transient process of electrical appliance by convolutional neural networks, and then identifies the type and state of electrical appliance;
S4 identifies the change procedure between electrical appliance stable state by C means Method, and then identifies the type and shape of electrical appliance State;
S5 determines the type and state of electrical appliance in conjunction with step S3 and S4.
On the basis of above technical scheme, it is preferred that the step S1 includes,
S1-1, samples in total on-line monitoring of electrical appliance, and the current signal denoised extracts fundamental current using sliding window Amplitude;
S1-2 carries out wave crest detection using the two dimensional image in step S1-1;
S1-3 carries out bilateral CUSUM detection using sliding window, by increasing between peak value to the two dimensional image in step S1-2 Variation reduces the accumulation of variation and the judgement with threshold value, judges whether change between the peak value of fundamental current in sliding window;
S1-4, according to threshold value, will be detected in step S1-3 the close sliding window section of the distance of fundamental current change procedure into Row fusion;
S1-5, the steady-state process before starting storage change after detecting variation in the sample of sampling, detection variation start Current signal afterwards, until until multiple samplings in variation is not detected.
It is further preferred that being the sliding window of a fundamental wave length using Period Length, through the past in the step S1-1 It is slided in sampled data after making an uproar, fast Fourier variation is done inside sliding window, extract the width of fundamental current in fundamental wave length Value, neglects the influence of harmonic current, then using fundamental voltage amplitude as ordinate, using the ordinal number of sliding window in sampled value as abscissa, Construct two dimensional image.
On the basis of above technical scheme, it is preferred that step S2 includes,
S2-1 carries out the current signal detected in step S1 after the sampling period of multiple variations is not detected Denoising, splicing obtain the curent change image of transient process and the characteristic value of the steady-state process on the way circuit of transient state front and back;
S2-2 compares whether the variation between the active power of the steady-state process on transient process both sides is more than threshold value, and judgement should Whether way circuit is added electrical appliance after transient process;If active power variation is more than threshold value, step S3 or S4 are executed, conversely, then Ignore this section of transient process, jumps back to step S1.
On the basis of above technical scheme, it is preferred that step S3 includes that the spliced electric current for obtaining step S2 is believed It number is added to based in the trained convolutional neural networks of electrical appliance transient current delta data, identifies the kind of possible electrical appliance Class and state.
On the basis of above technical scheme, it is preferred that step S4 includes,
S4-1 handles the steady state characteristic value before and after temporal variations that step S2 is obtained, according to obtaining front and back respectively The variable quantity of the corresponding fundamental current of two steady-state process, harmonic current and harmonic content generates feature vector;
Feature vector is input in the model of C mean cluster by S4-2, according in the currently known classification of vector distance The distance of the heart obtains possible electrical appliance type and state.
It is further preferred that the current amplitude variation characteristic value base in step S4-1, under each frequency in feature vector It is sought in following mathematic(al) representation:
Electric current under stable state can be decomposed into I=I1+I2+...+In
Wherein I1For fundamental current, InFor harmonic current, n ∈ [2, n], the electric current under each frequency can be by sine FunctionIt indicating, A is amplitude, and ω is angular frequency,For initial phase, the transient state that detected The knots modification of the electric current under each frequency between the stable state on both sides are as follows:
Wherein, IafterRepresent the electric current after transient state in the same frequency of stable state, IbeforeRepresent the same frequency of stable state after transient state Electric current in rate, IafterThe variable quantity for representing the electric current between stable state in same frequency, based on related sinusoidal and cosine formula, above formula It is alterable are as follows:
A in formulaafter, Abefore,It is right all it is known that the current amplitude knots modification under the frequency can be found out The electric current of each frequency seeks the knots modification of current amplitude, while can acquire the harmonic content in variation, generates feature vector.
On the basis of above technical scheme, it is preferred that the step S5 includes determining electrical appliance by the identification of step S3 Which classification belonged to, determines which brand electrical appliance belongs to by step S4.
It is further preferred that respectively by step S3 with the probability of the same category in step S4 multiplied by being added after weight, take it The mode of the maximum value of sum determines which classification electrical appliance belongs to, wherein weights sum 1, the classification in some step When not depositing, corresponding probability is 0.
The non-intruding load-type recognition methods of variation characteristic between synthesis transient characteristic of the invention and stable state is relative to existing There is technology to have the advantages that
(1) small to original data model dependence using C mean cluster, with the operation of system, used in continuous accumulation After the electrical appliance state of family home circuit changes feature vector, the feature database of the electrical appliance in training fitting subscriber household circuit, Improve the adaptability to domestic circuit environment of system.In this way, can not only prevent due to the electric appliance in electric appliance use process Aging, the change of caused electrical appliance state feature can also collect electrical appliance feature not of the same race, and the big data after helping is ground Study carefully electrical appliance of the same race.Meanwhile continue to optimize database, rather than it is simple data are stored in database, advantageously reduce data Expense needed for storage.
(2) transient state of electrical appliance is although changeable, but by research it can be found that electrical appliance of the same race have it is unique temporarily State feature.Electrical appliance transient event is identified using convolutional neural networks, increases having for a new dimension for identification electrical appliance Use information.Transient event and steady state characteristic are combined, accurately identifying to electrical appliance has help well.
(3) variable quantity between the stable state that the mathematical principle based on sinusoidal signal superposition extracts, than not considering in home circuit Phase angle, it is more accurate that the current amplitude of respective frequencies is simply subtracted each other to the feature vector changed.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with It obtains other drawings based on these drawings.
The non-intruding load-type recognition methods of variation characteristic of the Fig. 1 between synthesis transient characteristic and stable state of the invention it is temporary The data model training module flow chart that state, stable state change;
The stream of the non-intruding load-type recognition methods of variation characteristic of the Fig. 2 between synthesis transient characteristic and stable state of the invention Journey total figure;
The non-intruding load-type recognition methods of variation characteristic of the Fig. 3 between synthesis transient characteristic and stable state of the invention changes Into CUSUM event detection flow chart;
It is tied in the non-intruding load-type recognition methods of variation characteristic of the Fig. 4 between synthesis transient characteristic and stable state of the invention Close the identification electrical appliance flow chart of transient state and stable state variation characteristic;
Fig. 5 is C mean cluster model optimization flow chart;
The non-intruding load-type recognition methods of variation characteristic of the Fig. 6 between synthesis transient characteristic and stable state of the invention is System frame diagram.
Specific embodiment
Below in conjunction with embodiment of the present invention, the technical solution in embodiment of the present invention is carried out clearly and completely Description, it is clear that described embodiment is only some embodiments of the invention, rather than whole embodiments.Base Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts all Other embodiments shall fall within the protection scope of the present invention.
As shown in fig. 6, the non-intruding load-type identification side of the variation characteristic between synthesis transient characteristic of the invention and stable state The system framework figure of method, the data model including home circuit bus acquisition module, identification module, transient state, stable state variation are trained Module, storing data timing optimization module, home circuit electrical appliance lower state variation characteristic memory module and system and user Interactive module.Wherein, home circuit bus acquisition module will be mounted in the ammeter box of user, acquire the electricity of subscriber household circuit Sampled point is transferred to the event detection and identification that identification module carries out electrical appliance by the live signal of pressure and electric current.
Before system operation, it need to be trained in the way of Fig. 1 temporary between each state change for obtaining basic household electrical appliance The data characteristics of state and stable state variation.Specifically, include the following steps,
The electric current and voltage of sampled point when switching between the switch state and multimode of acquisition electrical appliance first;
Then, the stable state period and transient state period under each state of electrical appliance are distinguished;
For the transient state period, the current image in transient state is imported into convolutional neural networks training pattern, obtains transient process Electric current change convolutional neural networks model;
For the stable state period, by the stable state at transient state both ends, between being changed based on the state that sinusoidal Superposition Formula obtains electrical appliance Feature database obtain the C mean cluster model of the feature between each state change of electrical appliance further according to C mean cluster.
Since three major class: (1) resistive load, generally electrothermal load, active power can be divided into household electrical appliance Larger, start/stop impact is minimum, even without being represented as electric cooker, electric heating installation using oil as medium, insulating pot, heater;(2) inductive load, it is interior Portion has motor or inductance coil, and start and stop event often with wave forms impact, is represented as fluorescent lamp, electric fan, refrigerator, micro-wave oven etc.; (3) capacitive load, containing start and stop power supply, start and stop event often has surge waveform to be represented as LCD TV, computer.To each major class Load extracts 10 kinds of representational electrical appliances, and utilizes web crawlers major shopping and grading website from network, obtains every kind Electrical appliance 5 kinds of brands most popular with users acquire electricity when their single working conditions change to the electrical appliance of these brands Pressure and current data generate the convolutional Neural of the transient state of home circuit electrical appliance state change using the data training method of Fig. 1 The C mean cluster model changed between network model and stable state.
Wherein, the transient state when each state for distinguishing electrical appliance changes and the mode of steady-state process are taken different from CUSUM The incident detection method based on sliding window, using the library pandas and numpy of python, by the real-time voltage of sampling and electric current number According to the realtime power being calculated, according to the sliding window size N of setting, it is long that the instantaneous power in time domain is reconstructed into a row vector Degree is that the new matrix of N calculates its average value and variance to the row vector of matrix, if the variance in row vector is multiplied by a threshold Value alpha (recognition effect of alpha ∈ [0.0.5] is best in experiment) is more than the average value inside the row vector, then judges The row vector electric signal changes.After the operation, the sliding window by adjacent variation combines, and obtains the stable state of its signal Region of variation further obtains the stationary zones of sampled signal.This method the state of single electrical appliance is changed have than CUSUM is faster and more accurate recognition effect, but the state for being not suitable for the electrical appliance of a variety of high or low powers changes situation, because This is used for transient state and the stable state identification of single electrical appliance when training data.
Convolutional neural networks model is completed by caffee image-net training, and the data model of C mean cluster utilizes The packet of fuzz in python is quickly found out the ginseng of suitable C mean cluster data model according to training data using genetic algorithm Number, the i.e. number of cluster and the value of FPC, so that F-scores highest, that is, obtain optimal C mean cluster model.It can After completing training data, the C mean cluster model changed between the convolutional neural networks model and stable state of transient state is stored temporary In the data model training module that state, stable state change.
As shown in Fig. 2, the non-intruding load-type identification side of the variation characteristic between synthesis transient characteristic of the invention and stable state Method includes the following steps,
S1 takes monitoring storage mode, after fundamental current amplitude changes, starts to store the sampling period.To adopting Before sample signal carries out event detection, needs to carry out wavelet filtering to the electric current and voltage signal of sampling, filter out in sampled data Gaussian Background white noise and spike burr, the influence for preventing the noise on way circuit from identifying in load.
Specifically, as shown in figure 3, the step S1 includes,
S1-1, samples in total on-line monitoring of electrical appliance, and the current signal denoised extracts fundamental current using sliding window Amplitude.Specifically, being the sliding window of a fundamental wave length using Period Length, after denoising in the step S1-1 It is slided in sampled data, fast Fourier variation is done inside sliding window, extracted in fundamental wave length, the amplitude of fundamental current, suddenly The influence of harmonic current is omitted, then using fundamental voltage amplitude as ordinate, using the ordinal number of sliding window in sampled value as abscissa, building two Tie up image.
S1-2 carries out wave crest detection using the two dimensional image in step S1-1, only extracts these peak values reconstruct two dimensional image, Reduce the operand of CUSUM algorithm.
S1-3 carries out bilateral CUSUM detection using sliding window, by increasing between peak value to the two dimensional image in step S1-2 Variation reduces the accumulation of variation and the judgement with threshold value, judges whether change between the peak value of fundamental current in sliding window.Often Sliding window slides to the right half of sliding window length.
S1-4, according to threshold value, will be detected in step S1-3 the close sliding window section of the distance of fundamental current change procedure into Row fusion.
S1-5, the steady-state process before starting storage change after detecting variation in the sample of sampling, detection variation start Current signal afterwards, until until multiple samplings in variation is not detected.
S2 after the sampling period of multiple variations is not detected, then identifies the transient process and steady-state process of electrical appliance, needle Transient process to the electrical appliance identified executes step S3;For the steady-state process of the electrical appliance identified, step is executed S4。
Specifically, as shown in figure 4, step S2 includes,
S2-1 carries out the current signal detected in step S1 after the sampling period of multiple variations is not detected Denoising, splicing obtain the curent change image of transient process and the characteristic value of the steady-state process on the way circuit of transient state front and back. Characteristic value includes the phase angle between active power, voltage and current, and fundamental current amplitude, odd harmonic amplitude and harmonic wave contain Amount.The step is prevented since electrical appliance transient state time is too long, and the data of transient event are present in the data in multiple sampling periods In, caused transient event loss of data.
S2-2 compares whether the variation between the active power of the steady-state process on transient process both sides is more than threshold value, and judgement should Whether way circuit is added electrical appliance after transient process;If active power variation is more than threshold value, step S3 or S4 are executed, conversely, then Ignore this section of transient process, jumps back to step S1.
S3 is identified the transient process of electrical appliance by convolutional neural networks, and then identifies the type and state of electrical appliance.Tool Body, step S3, which includes that the spliced current signal for obtaining step S2 is added to, changes number based on electrical appliance transient current According to the type and state in trained convolutional neural networks, identifying possible electrical appliance.
S4 identifies the change procedure between electrical appliance stable state by C means Method, and then identifies the type and shape of electrical appliance State.Specifically, step S4 includes,
S4-1 handles the steady state characteristic value before and after temporal variations that step S2 is obtained, according to obtaining front and back respectively The variable quantity of the corresponding fundamental current of two steady-state process, harmonic current and harmonic content generates feature vector.Step S4-1 In, the current amplitude variation characteristic value under each frequency in feature vector is sought based on following mathematic(al) representation:
Electric current under stable state can be decomposed into I=I1+I2+...+In
Wherein I1For fundamental current, InFor harmonic current, n ∈ [2, n], the electric current under each frequency can be by sine FunctionIt indicating, A is amplitude, and ω is angular frequency,For initial phase, the transient state that detected The knots modification of the electric current under each frequency between the stable state on both sides are as follows:
Wherein, IafterRepresent the electric current after transient state in the same frequency of stable state, IbeforeRepresent the same frequency of stable state after transient state Electric current in rate, IafterRepresent the variable quantity of the electric current between stable state in same frequency.Based on related sinusoidal and cosine formula, above formula It is alterable are as follows:
Therefore,
A in formulaafter, Abefore,All it is known that the current amplitude knots modification under the frequency can be found out.According to According to this method, the knots modification of current amplitude is sought to the electric current of each frequency, while can acquire the harmonic content in variation, thus Generate feature vector.When the data model changed between the stable state of training different conditions, also sought under each frequency with the formula Current amplitude knots modification is characterized value.
Feature vector is input in the model of C mean cluster by S4-2, according in the currently known classification of vector distance The distance of the heart obtains possible electrical appliance type and state.
S5 determines the type and state of electrical appliance in conjunction with step S3 and S4.Specifically, the step S5 includes, by step The identification of S3 determines which classification electrical appliance belongs to, and determines which brand electrical appliance belongs to by step S4.Specifically, respectively The mode of the maximum value of its sum is taken to determine electrical appliance multiplied by being added after weight with the probability of the same category in step S4 step S3 Which classification belonged to, wherein weights sum 1, when the classification in some step is not deposited, corresponding probability is 0.
Specifically, being opened in time for a long time in the event detection of CUSUM according in view of the too small electrical appliance of power Too many electricity will not be consumed, the principle based on CUSUM, the middle threshold value for judging the event of changing that CUSUM is arranged is 0.14, I.e. in circuit, it when fundamental current amplitude accumulated change is more than 0.14A, and regards as changing.
According to experiment, it has also been found that, air-conditioning is the electrical appliance for possessing longest transient process in all electrical appliances, it is contemplated that The section working condition between being in when micro-wave oven work.It is set in being merged between change point according to the distance sliding window in order to prevent Be set to meet regard the transient process of air-conditioning as an entirety and cause micro-wave oven stable state detect in can not detect stable state Stable state, by training examples, given threshold is 20 cycle durations for we, i.e., will be less than threshold value when the distance between change point, i.e., It is merged, one whole section of transient process of acquisition.The electrical appliance of long-time transient state this for air-conditioning, experiment discovery is in transient state Portion there is also a certain section of fundamental current amplitude position is relatively stable, which is also regarded as be air-conditioning a duration compared with Short steady-state process, the change the transient-wave of experience required for reaching this stable state to air-conditioning and the stable state for reaching the stable state Change feature to be studied, constructs its feature database.
As shown in figure 5, in the eigenvector recognition electrical appliance of the current signal and stable state variation that rely on temporal variations In type, for the algorithm optimization of the C mean cluster in figure six after realizing, feature vector is being input to C mean cluster mould After being judged in type, will with (access time, electrical appliance label, stable state variation feature vector) storage beyond the clouds one individually In database, " change histories database between subscriber household electrical appliance stable state " is named as.
As shown in fig. 6, storing data library timing optimization module can carry out automatic in idle after system works one month System maintenance optimizes C mean cluster model.Contain following steps:
(1) firstly, home circuit electrical appliance feature extraction C mean cluster model contains storage system initial training model The database of feature vector (time of initial vector determines by the initialization time that system brings into operation, and it is poly- to be named as C mean value Category feature library), and become between the obtained 3 months subscriber household electrical appliance stable states of data by acquiring subscriber household circuit Change historical data base.When system timer therein reaches 3 months, reach flat for a long time with the active power of total current circuit Active power when working without other electrical appliances in normal home circuit is the standard of home circuit idle, and current home circuit reaches When to this standard, above-mentioned two databases are fused to train the characteristic vector data library of C mean cluster.According to 80% Tranining database and 20% test database selection random in fused database.
(2) in C mean cluster, the number for the cluster for updating point according to upper one initializes current cluster Number, obtain the center of new cluster after training for the parameter of C mean cluster training pattern.
(3) the correlation value between each class central feature vector is calculated, when class center vector similarity is high When up to 75%, two classes are merged, the increase of cluster is stopped, generating new C mean cluster model, and by current class In feature vector preceding 500 feature vectors that behave oneself best be deposited into C characteristics of mean database.The performance of feature vector is by depositing The distance at the angle of incidence and distance-like center determines.The access time of feature vector in C characteristics of mean database is unified for The time of model optimization, meanwhile, the data of change histories database will be emptied between subscriber household electrical appliance stable state.If between class Correlation gap is still very big, then needs to continue growing the number of cluster, carries out C mean cluster, and repeat step (2) process in.
The foregoing is merely better embodiments of the invention, are not intended to limit the invention, all of the invention Within spirit and principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.

Claims (9)

1. a kind of non-intruding load-type recognition methods of the variation characteristic between comprehensive transient characteristic and stable state, it is characterised in that: packet Include following steps,
S1 takes monitoring storage mode, after fundamental current amplitude changes, starts to store the sampling period;
S2 after the sampling period of multiple variations is not detected, then identifies the transient process and steady-state process of electrical appliance, for knowledge Not Chu electrical appliance transient process, execute step S3;For the steady-state process of the electrical appliance identified, step S4 is executed;
S3 is identified the transient process of electrical appliance by convolutional neural networks, and then identifies the type and state of electrical appliance;
S4 identifies the change procedure between electrical appliance stable state by C means Method, and then identifies the type and state of electrical appliance;
S5 determines the type and state of electrical appliance in conjunction with step S3 and S4.
2. the non-intruding load-type recognition methods of the variation characteristic between comprehensive transient characteristic and stable state as described in claim 1, It is characterized by: the step S1 includes,
S1-1 is sampled in total on-line monitoring of electrical appliance, and the current signal denoised extracts the width of fundamental current using sliding window Value;
S1-2 carries out wave crest detection using the two dimensional image in step S1-1;
S1-3 carries out bilateral CUSUM detection using sliding window, is changed by increasing between peak value to the two dimensional image in step S1-2 Or the accumulation of variation and the judgement with threshold value are reduced, judge whether change between the peak value of fundamental current in sliding window;
S1-4 will detect that the close sliding window section of the distance of fundamental current change procedure is melted according to threshold value in step S1-3 It closes;
S1-5, the steady-state process before starting storage change after detecting variation in the sample of sampling, after detection variation starts Current signal, until until multiple samplings in variation is not detected.
3. the non-intruding load-type recognition methods of the variation characteristic between comprehensive transient characteristic and stable state as claimed in claim 2, It is characterized by: being the sliding window of a fundamental wave length using Period Length, in the sampling after denoising in the step S1-1 It is slided in data, fast Fourier variation is done inside sliding window, extracted in fundamental wave length, the amplitude of fundamental current neglects The influence of harmonic current, using the ordinal number of sliding window in sampled value as abscissa, constructs X-Y scheme then using fundamental voltage amplitude as ordinate Picture.
4. the non-intruding load-type recognition methods of the variation characteristic between comprehensive transient characteristic and stable state as described in claim 1, It is characterized by: step S2 includes,
S2-1, after the sampling period of multiple variations is not detected, the current signal detected in step S1 is denoised, Splicing obtains the curent change image of transient process and the characteristic value of the steady-state process on the way circuit of transient state front and back;
S2-2 compares whether the variation between the active power of the steady-state process on transient process both sides is more than threshold value, judges the transient state Whether way circuit is added electrical appliance after process;If active power variation is more than threshold value, step S3 or S4 are executed, conversely, then ignoring This section of transient process jumps back to step S1.
5. the non-intruding load-type recognition methods of the variation characteristic between comprehensive transient characteristic and stable state as described in claim 1, It is characterized by: step S3 includes, the spliced current signal that step S2 is obtained is added to based on electrical appliance transient current In the trained convolutional neural networks of delta data, the type and state of possible electrical appliance are identified.
6. the non-intruding load-type recognition methods of the variation characteristic between comprehensive transient characteristic and stable state as described in claim 1, It is characterized by: step S4 includes,
S4-1 handles the steady state characteristic value before and after temporal variations that step S2 is obtained, according to obtaining former and later two respectively The variable quantity of the corresponding fundamental current of steady-state process, harmonic current and harmonic content generates feature vector;
Feature vector is input in the model of C mean cluster by S4-2, according to the center of the currently known classification of vector distance Distance obtains possible electrical appliance type and state.
7. the non-intruding load-type recognition methods of the variation characteristic between comprehensive transient characteristic and stable state as claimed in claim 6, It is characterized by: the current amplitude variation characteristic value under each frequency in feature vector is based on following number in step S4-1 Expression formula is learned to seek:
Electric current under stable state can be decomposed into I=I1+I2+...+In
Wherein I1For fundamental current, InFor harmonic current, n ∈ [2, n], the electric current under each frequency can be by SIN functionIt indicating, A is amplitude, and ω is angular frequency,For initial phase, the transient state both sides that detected Stable state between each frequency under electric current knots modification are as follows:
Wherein, IafterRepresent the electric current after transient state in the same frequency of stable state, IbeforeIt represents after transient state in the same frequency of stable state Electric current, IafterThe variable quantity for representing the electric current between stable state in same frequency, based on related sinusoidal and cosine formula, above formula is variable It turns to:
A in formulaafter, Abefore,All it is known that the current amplitude knots modification under the frequency can be found out, to each frequency The electric current of rate seeks the knots modification of current amplitude, while can acquire the harmonic content in variation, generates feature vector.
8. the non-intruding load-type recognition methods of the variation characteristic between comprehensive transient characteristic and stable state as described in claim 1, It is characterized by: the step S5 includes, determines which classification electrical appliance belongs to by the identification of step S3, determined by step S4 Which brand electrical appliance belongs to.
9. the non-intruding load-type recognition methods of the variation characteristic between comprehensive transient characteristic and stable state as claimed in claim 8, It is characterized by: respectively by step S3 with the probability of the same category in step S4 multiplied by being added after weight, take the maximum value of its sum Mode determine which classification electrical appliance belongs to, wherein weights sum 1, it is corresponding when the classification in some step is not deposited Probability be 0.
CN201811594124.1A 2018-12-25 2018-12-25 Non-intrusive load type identification method integrating transient characteristic and change characteristic between stable states Active CN109813978B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811594124.1A CN109813978B (en) 2018-12-25 2018-12-25 Non-intrusive load type identification method integrating transient characteristic and change characteristic between stable states

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811594124.1A CN109813978B (en) 2018-12-25 2018-12-25 Non-intrusive load type identification method integrating transient characteristic and change characteristic between stable states

Publications (2)

Publication Number Publication Date
CN109813978A true CN109813978A (en) 2019-05-28
CN109813978B CN109813978B (en) 2021-04-20

Family

ID=66602331

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811594124.1A Active CN109813978B (en) 2018-12-25 2018-12-25 Non-intrusive load type identification method integrating transient characteristic and change characteristic between stable states

Country Status (1)

Country Link
CN (1) CN109813978B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110456130A (en) * 2019-07-24 2019-11-15 清华大学 A method of transient voltage measured waveform is decoupled using three-phase steady state voltage information
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

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103020459A (en) * 2012-12-19 2013-04-03 中国科学院计算技术研究所 Method and system for sensing multiple-dimension electric utilization activities
CN103105556A (en) * 2013-01-30 2013-05-15 西安交通大学 Intelligent power grid load testing and recognition method based on steady state and transient state characteristic joint matching
CN105162162A (en) * 2015-09-07 2015-12-16 天津大学 Phase perturbation based island detection system for distributed gird-connected inverter
CN106655160A (en) * 2016-10-27 2017-05-10 国家电网公司 Non-intrusion electric power load decomposition identification decision method and system
CN108459222A (en) * 2018-03-22 2018-08-28 中国海洋大学 A kind of electric appliance fault detection method and system
CN108899892A (en) * 2018-06-08 2018-11-27 上海电力学院 One kind being based on CNN non-intrusion type power load decomposition method

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103020459A (en) * 2012-12-19 2013-04-03 中国科学院计算技术研究所 Method and system for sensing multiple-dimension electric utilization activities
CN103105556A (en) * 2013-01-30 2013-05-15 西安交通大学 Intelligent power grid load testing and recognition method based on steady state and transient state characteristic joint matching
CN105162162A (en) * 2015-09-07 2015-12-16 天津大学 Phase perturbation based island detection system for distributed gird-connected inverter
CN106655160A (en) * 2016-10-27 2017-05-10 国家电网公司 Non-intrusion electric power load decomposition identification decision method and system
CN108459222A (en) * 2018-03-22 2018-08-28 中国海洋大学 A kind of electric appliance fault detection method and system
CN108899892A (en) * 2018-06-08 2018-11-27 上海电力学院 One kind being based on CNN non-intrusion type power load decomposition method

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
CN110456130A (en) * 2019-07-24 2019-11-15 清华大学 A method of transient voltage measured waveform is decoupled using three-phase steady state voltage information
CN110456130B (en) * 2019-07-24 2020-10-16 清华大学 Method for decoupling transient voltage measurement waveform by using three-phase steady-state voltage information

Also Published As

Publication number Publication date
CN109813978B (en) 2021-04-20

Similar Documents

Publication Publication Date Title
Smith et al. Demand response: a strategy to address residential air-conditioning peak load in Australia
Wong et al. Recent approaches to non-intrusive load monitoring techniques in residential settings
CN109813978A (en) A kind of non-intruding load-type recognition methods of variation characteristic between comprehensive transient characteristic and stable state
Ducange et al. A novel approach based on finite-state machines with fuzzy transitions for nonintrusive home appliance monitoring
Benetti et al. Electric load management approaches for peak load reduction: A systematic literature review and state of the art
Ahmadi et al. Load decomposition at smart meters level using eigenloads approach
CN103135009B (en) Based on electrical equipment method for detecting and the system of user's feedback information
CN106096726A (en) A kind of non-intrusion type load monitoring method and device
CN106815677A (en) The recognition methods of non-intrusion type load and device
EP2026299A1 (en) Cognitive electric power meter
CN106936129A (en) Electric load discrimination method based on multi-feature fusion and system
CN109325545A (en) Low-voltage network topological structure method of calibration, device, equipment and storage medium
CN109697548A (en) Electrical energy consumption analysis server and its electrical energy consumption analysis method
CN108021736B (en) Load switching action monitoring method based on sliding window residual error model
Najafi et al. Data analytics for energy disaggregation: methods and applications
Batra et al. Neighbourhood nilm: A big-data approach to household energy disaggregation
CN109470957B (en) Non-invasive electrical appliance identification method based on Frey's distance
Naghibi et al. Non-intrusive load monitoring and supplementary techniques for home energy management
Selvam et al. Fuzzy based clustering of smart meter data using real power and THD patterns
CN108767847B (en) Non-invasive electric load decomposition method and device
CN110321934A (en) A kind of method and system detecting user power utilization abnormal data
CN109541346A (en) A kind of non-intrusion type electrical load under steady state condition identifies method for improving
Chen et al. A novel system for extracting useful correlation in smart home environment
Khazaei et al. Trends and challenges in smart metering analytics
Veleva et al. Box-dimension as a correlation measure for data mining of power socket sensor data

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
GR01 Patent grant
GR01 Patent grant