CN103020459B - A kind of cognitive method of various dimensions electricity consumption behavior and system - Google Patents

A kind of cognitive method of various dimensions electricity consumption behavior and system Download PDF

Info

Publication number
CN103020459B
CN103020459B CN201210555210.8A CN201210555210A CN103020459B CN 103020459 B CN103020459 B CN 103020459B CN 201210555210 A CN201210555210 A CN 201210555210A CN 103020459 B CN103020459 B CN 103020459B
Authority
CN
China
Prior art keywords
electricity consumption
current
consumption behavior
family
weight
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.)
Active
Application number
CN201210555210.8A
Other languages
Chinese (zh)
Other versions
CN103020459A (en
Inventor
刘晶杰
徐志伟
聂磊
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.)
Institute of Computing Technology of CAS
Original Assignee
Institute of Computing Technology of CAS
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 Institute of Computing Technology of CAS filed Critical Institute of Computing Technology of CAS
Priority to CN201210555210.8A priority Critical patent/CN103020459B/en
Publication of CN103020459A publication Critical patent/CN103020459A/en
Application granted granted Critical
Publication of CN103020459B publication Critical patent/CN103020459B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Abstract

The invention provides a kind of cognitive method and system of various dimensions electricity consumption behavior, the method comprises: step 1, is previously defined in the statistical nature of consumer under multiple dimension and the eigenwert under obtaining often kind of statistical nature, and sets up corresponding statistical model; Step 2, installs collecting unit in electrical network porch, the current-voltage information of all consumers of Real-time Obtaining, and computing unit trains the feature weight of each statistical nature according to the input Data Comparison of described current-voltage information and user; Step 3, described computing unit carries out heuristic search calculating according to described eigenwert, statistical model and feature weight, obtains the real-time electricity consumption behavior of each consumer, and described real-time electricity consumption behavior is sent to display unit.The present invention uses multiple load characteristic to combine series of computation method statistically and directly obtains electricity consumption behavioural information in family on all devices, and resolution characteristic is powerful, judge that precision is high, effectively can distinguish the electricity consumption behavior of similar consumer.

Description

A kind of cognitive method of various dimensions electricity consumption behavior and system
Technical field
The present invention relates to Computer Applied Technology field, the cognitive method of particularly a kind of various dimensions electricity consumption behavior and system.
Background technology
IBM in 2008 proposes the concept of " the wisdom earth ", is described as by the wisdom earth " more thorough perception, more fully interconnect, more deep intellectuality ".U.S. government proposes in the near future also to be brought into schedule as the pith in its " plan of new forms of energy rescue market " by intelligent grid at the wisdom earth.By comparison, intelligent grid means and obtains more information as far as possible the traditional electrical network known with us, and that more pays attention between user is mutual, is provided the service more put in place by resolving information.The ustomer premises access equipments such as intelligent electric meter directly contact with user, instruct user power utilization behavior, are the important embodiments that intelligent grid is different from traditional electrical network.By providing the detailed electricity consumption situation in relevant devices, effectively can reduce the understanding deviation of user to electricity consumption behavior, the use habit of optimizing user, thus obtaining better power savings.
Therefore, effectively obtaining the detailed power information in relevant devices in power utilization environment (family, production environment etc.), is the gordian technique that intelligent grid field user client information gathers.When not affecting power utilization environment, obtaining the monitoring technology of the detailed power information each equipment from outside, being called as non-intrusion type load monitoring (NILM) technology.Up to the present, non-intrusion type load monitoring technology mainly comprises two large classes: the load monitoring technology based on steady-state analysis and the load monitoring technology based on transient event.Although these two kinds of technology all support the load monitoring of non-intrusion type, meet the demand that intelligent grid field user client information gathers, but these two kinds of technology have all made similar hypothesis to consumer: equipment has metastable running status, can according to known information after determining running status, the detailed power information of equipment.
Along with the epoch are progressive, the behavior elastification day by day of consumer, makes this hypothesis no longer applicable, such as: in the same time period, the computer of running game program is relative to browsing merely the more electric power of webpage consumption, and idle with busy power consumption difference may more than 30%.Meanwhile, the two dimension electrical feature of traditional active power, reactive power composition is not enough to a large amount of similar electrical equipment distinguishing current use.Even if introduce higher hamonic wave to expand characteristic set, power supply adaptor similar in a large number still can produce the harmonic signal easily obscured.
Summary of the invention
The present invention is in the thinking of traditional steady-state analysis, multiple load characteristic is used to redefine the steady state (SS) of consumer, analyze in conjunction with the whole family front yard power information of series of computation method statistically to Real-time Collection, directly obtain the electricity consumption behavioural information in family on all devices, resolution characteristic is powerful, judge that precision is high, effectively can distinguish the electricity consumption behavior of similar consumer.
For achieving the above object, the invention provides a kind of cognitive method of various dimensions electricity consumption behavior, comprising:
Step 1, is previously defined in the statistical nature of consumer under multiple dimension and the eigenwert under obtaining often kind of statistical nature, and sets up corresponding statistical model;
Step 2, installs collecting unit in electrical network porch, the current-voltage information of all consumers of Real-time Obtaining, and computing unit, according to the input data of described current-voltage information and user, is trained the feature weight of each statistical nature;
Step 3, described computing unit carries out heuristic search calculating according to described eigenwert, statistical model and feature weight, obtains the real-time electricity consumption behavior of each consumer, and described real-time electricity consumption behavior is sent to display unit.
Further, described collecting unit comprises a current sensor and a voltage sensor, this current sensor is the magnetic field induction chip based on Hall effect, electromagnetic induction is utilized to realize the current data collection of non-intrusion type, the direct place in circuit of this voltage sensor is in parallel with all consumers, measures the voltage on it.
Described computing unit independently uses a processor; Or share a processor with collecting unit; Or share a processor with display unit.
Further, described step 2 comprises:
Step 21, measures the total current information of voltage of all consumers in the family in a period of time, starts to carry out weight training as training dataset,
Step 22, is calculated by described heuristic search, obtains the candidate list of multiple candidate result composition;
Step 23, user chooses optimum analysis result according to truth in candidate list;
Step 24, carries out iterative optimization weight parameter by the information provided described candidate list and user, obtains the feature weight of described all consumers.
Further, the heuristic computing formula in described step 3 is:
P ( c | t ) = exp Σ i = 1 n λ i h i ( c , t ) Σ c ′ exp Σ i = 1 n λ i h i ( c ′ , t )
Wherein, t represents character representation total in family, and c is the one electricity condition of all electrical equipment in family, and P (c|t) represents when known t, and in family, electricity consumption situation is just the probability of c; The h on right side ibe the fundamental function of a dimension, λ ifor the h of correspondence iweight parameter, denominator is by the summation of the probability of possible state c ' all in the family when t.
For achieving the above object, present invention also offers a kind of sensory perceptual system of various dimensions electricity consumption behavior, comprising:
Pretreatment module, is previously defined in the statistical nature of consumer under multiple dimension and the eigenwert under obtaining often kind of statistical nature, and sets up corresponding statistical model;
Training module, installs collecting unit in electrical network porch, the current-voltage information of all consumers of Real-time Obtaining, and computing unit trains the feature weight of each statistical nature according to the input Data Comparison of described current-voltage information and user
Computing module, described computing unit carries out heuristic search calculating according to described eigenwert, statistical model and feature weight, obtains the real-time electricity consumption behavior of each consumer, and described real-time electricity consumption behavior is sent to display unit.
Described collecting unit comprises a current sensor and a voltage sensor, this current sensor is the magnetic field induction chip based on Hall effect, electromagnetic induction is utilized to realize the current data collection of non-intrusion type, the direct place in circuit of this voltage sensor is in parallel with all consumers, measures the voltage on it.
Described computing unit independently uses a processor; Or share a processor with collecting unit; Or share a processor with display unit.
Described training module comprises:
Weight training preparation module, measures the current-voltage information of each consumer in a period of time, starts to carry out weight training as training dataset,
Candidate collection acquisition module, is calculated by described heuristic search, obtains the candidate list of multiple candidate result composition;
Choose module, user chooses optimum analysis result according to truth in candidate list;
Weight obtains module, carries out iterative optimization weight parameter, obtain the feature weight of described all consumers by the information provided described candidate list and user.
Heuristic computing formula in described computing module is:
P ( c | t ) = exp Σ i = 1 n λ i h i ( c , t ) Σ c ′ exp Σ i = 1 n λ i h i ( c ′ , t )
Wherein, t represents character representation total in family, and c is the one electricity condition of all electrical equipment in family, and P (c|t) represents when known t, and in family, electricity consumption situation is just the probability of c; The h on right side ibe the fundamental function of a dimension, λ ifor the h of correspondence iweight parameter, denominator is by the summation of the probability of possible state c ' all in the family when t.
Technique effect:
1) the present invention uses and is described equipment electricity condition from the statistical nature model possessing physical significance of different dimensions in a large number, and it organically can be combined, there is provided powerful resolution characteristic: existing non-intrusion type load monitoring technology uses same class physical quantity to carry out the description of equipment state mostly, afterwards again according to actual electricity consumption situation identification electrical equipment with the conversion between electricity condition or state, complete the analysis of power information.These class methods have a direct defect, and when equipment multiple in family has similar physical features, analysis precision can significantly decrease.And the present invention sets up the statistical nature of multiple different dimensions to equipment electricity consumption behavior and is combined into unified computation model, analytic process does not rely on the use electrical feature of single dimension, effectively can distinguish the electricity consumption behavior that like device produces.
2) the present invention is in electricity consumption behavior perception, employ the computing method of heuristic search, effectively in conjunction with the statistical nature model of various dimensions, very big raising electricity consumption condition adjudgement precision: existing non-intrusion type load monitoring technology is when carrying out equipment electricity consumption behavioural analysis, and what only can use single dimension uses electrical feature.And the present invention is under the unified quantization that various dimensions statistical nature is formed represents, uses heuristic search Direct Analysis equipment electricity condition, effectively follow the trail of the electricity consumption Behavioral change situation of each equipment, the mean accuracy of power information analysis can be brought up to 95%.
Describe the present invention below in conjunction with the drawings and specific embodiments, but not as a limitation of the invention.
Accompanying drawing explanation
Fig. 1 is the cognitive method process flow diagram of various dimensions electricity consumption behavior of the present invention;
Fig. 2 is electricity consumption behavior sensory perceptual system algorithm flow schematic diagram;
Fig. 3 is feature set establishment stage process flow diagram;
Fig. 4 is parameter learning phase flow figure;
Fig. 5 is behavioural analysis phase flow figure;
Fig. 6 is statistical nature environment schematic to be measured;
Fig. 7 is the electricity consumption behavior sensory perceptual system schematic diagram of household internal power utilization environment;
Fig. 8 is the sensory perceptual system schematic diagram of various dimensions electricity consumption behavior of the present invention.
Embodiment
Fig. 1 is the cognitive method process flow diagram of various dimensions electricity consumption behavior of the present invention.As shown in Figure 1, the method comprises:
Step 1, is previously defined in the statistical nature of consumer under multiple dimension and the eigenwert under obtaining often kind of statistical nature, and sets up corresponding statistical model;
Step 2, install collecting unit in electrical network porch, the current-voltage information of all consumers of Real-time Obtaining, computing unit is trained according to the feature weight of input data to each statistical nature of described current-voltage information and user;
Step 3, described computing unit carries out heuristic search calculating according to described eigenwert, statistical model and feature weight, obtains the real-time electricity consumption behavior of each consumer, and described real-time electricity consumption behavior is sent to display unit.
Further, described step 2 comprises:
Step 21, measures the total current information of voltage of all consumers in the family in a period of time, starts to carry out weight training as training dataset,
Step 22, by the heuristic search identical with step 3, obtains the candidate list of multiple candidate result composition;
Step 23, user chooses optimum analysis result according to truth in candidate list;
Step 24, carries out iterative optimization weight parameter by the information provided described candidate list and user, obtains the feature weight of described all consumers.
Further, the heuristic computing formula in described step 3 is:
P ( c | t ) = exp Σ i = 1 n λ i h i ( c , t ) Σ c ′ exp Σ i = 1 n λ i h i ( c ′ , t )
Wherein, t represents character representation total in family, and c is the one electricity condition of all electrical equipment in family, and P (c|t) represents when known t, and in family, electricity consumption situation is just the probability of c; The h on right side ibe the fundamental function of a dimension, λ ifor the h of correspondence iweight parameter, denominator is by the summation of the probability of possible state c ' all in the family when t.
Fig. 8 is the sensory perceptual system schematic diagram of various dimensions electricity consumption behavior of the present invention.As shown in Figure 8, this system comprises:
Pretreatment module 100, is previously defined in the statistical nature of consumer under multiple dimension and the eigenwert under obtaining often kind of statistical nature, and sets up corresponding statistical model;
Training module 200, install collecting unit in electrical network porch, the current-voltage information of all consumers of Real-time Obtaining, computing unit is trained according to the feature weight of input data to each statistical nature of described current-voltage information and user
Computing module 300, described computing unit carries out heuristic search calculating according to described eigenwert, statistical model and feature weight, obtains the real-time electricity consumption behavior of each consumer, and described real-time electricity consumption behavior is sent to display unit.
Described training module 200 comprises:
Weight training preparation module, measures the current-voltage information of each consumer in a period of time, starts to carry out weight training as training dataset,
Candidate collection acquisition module, is calculated by described heuristic search, obtains the candidate list of multiple candidate result composition;
Choose module, user chooses optimum analysis result according to truth in candidate list;
Weight obtains module, carries out iterative optimization weight parameter, obtain the feature weight of described all consumers by the information provided described candidate list and user.
Heuristic computing formula in described computing module 300 is:
P ( c | t ) = exp Σ i = 1 n λ i h i ( c , t ) Σ c ′ exp Σ i = 1 n λ i h i ( c ′ , t )
Wherein, t represents character representation total in family, and c is the one electricity condition of all electrical equipment in family, and P (c|t) represents when known t, and in family, electricity consumption situation is just the probability of c; The h on right side ibe the fundamental function of a dimension, λ ifor the h of correspondence iweight parameter.
Wherein collecting unit is main data input device, and electricity consumption behavior sensory perceptual system, by installing collecting unit near domestic electric network entrance, obtains the current-voltage information in whole family.Current sensor in collecting unit can use the magnetic field induction chip based on Hall effect, utilizes the principle of electromagnetic induction to realize the current data collection of non-intrusion type.And voltage sensor directly access mains circuit can be in parallel with all consumers, measure the voltage on it.Use in electrical feature what set up, frequency domain character has requirement for the frequency of image data, in one embodiment of the invention: during the eigenwert that algorithm uses Short Time Fourier Transform to calculate in domain space, require that each cycle comprises 256 sampled points, namely the alternating current of 50Hz is needed to the sample frequency of 12.8kHz.
Computing unit is the core of whole electrical energy consumption analysis system, and first the current-voltage information of whole family will be converted into various dimensions eigenwert (observing the set of the eigenwert in dimension in difference) herein, and then analyze the real-time power information obtaining each equipment.Computing unit obtains the circuit data of Real-time Collection from collecting unit, after calculating, analysis result is delivered to display unit.In an embodiment of the present invention, the equipment possessing arbitrarily enough computing powers can as computing unit in system.Therefore computing unit can independently realize, and also can share a processor with collecting unit, form the novel collecting device of similar intelligent electric meter; A processor can also be shared with display unit, form the novel display device of similar intelligent terminal.
Display unit is the formant of system and user interactions, except basic display and interactive function, also needs to carry out adding up to real-time power information and to analyze, and the storage and management function of supported data.Display unit gets the real-time power information of each consumer in family from computing unit, upgrades, and these contents be saved in database or other media the history power information of each equipment and whole family and real-time power information; On the other hand, display unit needs to realize a set of effective User Interface, comprises patterned display interface, by the form that power information can be understood to be converted into user, feeds back to user; User to analyzing the partial feedback of distortion to electricity consumption behavior sensory perceptual system, thus can improve with the precision of post analysis simultaneously.
Technological core of the present invention concentrates in the behavior perception algorithm that computing unit uses.Fig. 2-5 is process flow diagrams of behavior perception algorithm in the present invention.Algorithm flow can be divided into three phases (as Fig. 2): feature set establishment stage, parameter learning stage, behavioural analysis stage.
Feature set establishment stage is for each consumer sets up necessary stage of various dimensions statistical nature model.This stage starts, first according to the measurement capability of collecting unit, to be set in characteristic set required in whole computation process.In this one-phase, first gather electric current when individual equipment normally works and voltage signal; Then the data collected are converted to the various dimensions eigenwert preset; Corresponding statistical models is set up according to these eigenwerts.Statistical model can be set up respectively for each running status for the equipment (such as: the electrical equipment that under the different conditions such as air-conditioning, micro-wave oven, electricity consumption situation difference is larger) with multiple running status.This stage is completed by equipment supplier or third party professional institution usually, can obtain more accurate result.The eigenwert of each equipment obtained in this stage will be stored into database, use in the stage below.
The parameter learning stage is the weight of consumer training for balancing each feature.This stage should complete in subscriber household.In this one-phase, user normally uses the electrical equipment in family, the electric current produced at domestic electric network entrance when the multiple equipment of system acquisition normally works and voltage signal; Use initial weight (default value preset) to carry out electrical energy consumption analysis, obtain the candidate list of multiple candidate result composition; By the input comparison with user, iterative optimization weight parameter, realizes error rate and minimizes training.Each feature weight obtained in this stage will be stored into database, use in the next stage.
The behavioural analysis stage is the Main Stage of carrying out equipment electricity consumption behavioural analysis.In this one-phase, first need feature and the parameter thereof of from database, taking out all known devices in family, and choose suitable statistical analysis technique according to the structure of characteristic set; Then gather electric current and the voltage signal of whole family in electrical network porch, need the data collected to be converted to predefined eigenwert simultaneously; When known total power information, by the computing method of heuristic search, infer use electricity condition most possible on each equipment in current whole family; Obtain power information actual on each equipment thus, these information transmission carry out last process to display unit the most at last.
Introduce various dimensions electricity consumption behavior sensory perceptual system below in detail, introduce three Main Stage in the present embodiment in conjunction with process flow diagram respectively:
Feature set establishment stage: first need when this stage starts to define the characteristic set used, the characteristic set used in the present embodiment is the inherent feature of electrical equipment itself, so that feature can be reused in different families, comprising:
Harmonic current feature: the independent probability distribution of each order harmonic current value that equipment produces, is described the value of the harmonic current of equipment;
Correlative character between each order harmonics: joint probability distribution when equipment runs between different order harmonic electric current, is described the relation between the harmonic current of equipment;
Conductance feature: the conductance probability distribution of equipment when using equipment to run, is described the characteristic of the physical circuit of device interior.
In this stage, equipment supplier completes and measures the feature of each equipment.After equipment supplier starts feature set establishment stage, as shown in Figure 3 first according to the measurement capability of collecting unit, initialization needs the characteristic set set up.Then according to characteristic set defined above, the collecting device disposing enough accuracy completes building of test environment, namely installs Devices to test, deployed environment.In the environment, Devices to test is run, the current and voltage data in recording unit operational process.For the complex apparatus with different running status, need its all running status to measure respectively.Therefore need to judge whether to there is unmeasured running status, if there is unmeasured running status, then equipment is adjusted to steady operation under this running status, start image data simultaneously, according to image data, the characteristic set of computing equipment, repeats deterministic process, until there is not unmeasured running status, then preserve the characteristic set of all acquisitions.So far feature set establishment stage terminates.
After measurement completes, according to characterizing definition before, by original current and voltage data, calculate the statistical model of character pair.In the present embodiment, the process calculating harmonic current feature is: carry out discrete Fourier transformation to raw current data, obtain each harmonic.Distribution fitting method is used to calculate its probability distribution respectively to the harmonic wave of different order.This distribution is exactly the foundation calculating harmonic current feature.Any time, harmonic current feature was exactly the long-pending logarithm value that each runs the parameter probability valuing of electrical equipment.
Log-linear model is used for the statistical nature (h) of each dimension, is combined by the weight parameter (λ) of feature, the statistical probability with electricity condition that final acquisition is most possible, and its formula is as follows:
P ( c | t ) = exp Σ i = 1 n λ i h i ( c , t ) Σ c ′ exp Σ i = 1 n λ i h i ( c ′ , t )
Wherein, t represents character representation total in family, and c is the one electricity condition of all electrical equipment in family.P (c|t) represents when known t, and in family, electricity consumption situation is just the probability of c; The h on right side ibe the fundamental function of a dimension, λ ifor the h of correspondence iweight parameter, denominator is by the summation of the probability of possible state c ' all in the family when t.The physical significance of this formula is when overall electricity consumption situation, and in family, electricity consumption situation is just for the probability of particular state is proportional to the linear sum of the logarithm value of each multidimensional characteristic, is therefore called log-linear model, namely sets up the process of statistical nature model.
The parameter learning stage (as shown in Figure 4): this one-phase in this stage needs to complete in subscriber household, first user needs the statistical nature model obtaining each consumer from equipment supplier, imported sensory perceptual system, namely be loaded into the basic configuration situation of consumer in family, be loaded into the statistical nature model under each state of each equipment.System normally uses the total current of (not necessarily simultaneously working) and voltage signal as training dataset using gathering in family all devices in a period of time, then needs the character representation data collected being converted to the various dimensions preset.By initial weight, carry out weight training.System is from the general Initial parameter sets preset, calculated the candidate state set of multiple optimum by the searching method (being described in detail in the next stage) similar with the behavioural analysis stage, the set sizes of the present embodiment setting is 100.Namely to each moment, by present weight, the candidate state set that acquisition 100 is optimum is calculated.After calculating candidate collection first, submit to user, allow it select an optimum situation.Afterwards, according to the selection of user, the weight choosing certain feature adjusts, thus makes candidate result list trend towards the truth of user's input.Namely the process of iterative optimization characteristic parameter is started.This process chooses a feature at every turn, and the parameter lambda of this feature being regarded as is independent variable, by changing this parameter, can change the order of whole candidate collection, and the value of adjustment parameter, makes the sequence of candidate collection according to inputting identical as much as possible with user.Choose each feature successively afterwards, carry out iteration optimization.Complete one take turns tuning after, analyze quality whether improve, if improve, then returning right retraining step, until quality no longer improves.Finally all weight parameter of each equipment obtained will be stored into database, record current feature weight, as the optimal weights in family, to use in the next stage, the parameter learning stage terminates.
The behavioural analysis stage (as shown in Figure 5): statistical nature model and the weight parameter thereof of taking out all devices in family from database; Open collecting device, obtain the real-time electricity consumption data in family; According to the electricity consumption characteristic set set up in advance, the data collected are turned to characteristic of correspondence and represents; In conjunction with feature and the weight information of known each equipment, calculate the most probable electricity consumption situation of each equipment current by didactic post search; Result imports display unit into, completes the post-processed such as visual, statistics; Judge whether to receive halt command, if do not received, then repeat above-mentioned steps until receive halt command, the behavioural analysis stage terminates.In the stage, the character representation of known population, the method for being searched for by heuristic post calculates electricity consumption situation most possible on each equipment in whole family.The concrete grammar of the method for heuristic post search is as follows: first with the starting point of whole appliance system running status for search, and this state can choose the last result phase analyzed, and also can choose the state that all electrical equipment all cuts out; First the value of the log-linear model of original state is calculated; Then from initial system state, calculate each value corresponding with the running status that original state only has an electrical equipment to be in the system state of different operating state, only retain top n state (N is a constant of post search setting).Every one deck of calculating search tree like this, until the characteristic model value in new level does not improve.In all states searched, what characteristic model value was the highest is exactly the most possible duty of each equipment.Derive more detailed power information by the electricity consumption situation of each equipment obtained, final step these information transmission is carried out visualization processing to display unit and carries out statistical study.
Here is a specific embodiment of the present invention:
One. set up characteristic set, as follows:
Current harmonics feature: the independent probability distribution of each order harmonic current value that equipment produces;
The eigenwert of input is the harmonic current of all orders of all electrical equipment;
Output is the logarithm value of a statistical probability, the probability that the eigenwert just inputted therewith for all electrical equipment is identical.
Formula is:
h 1 ( { c ij } ) = log ( Π i = 1 n Π j = 1 m P ij ( c ij ) ) = Σ i = 1 n Σ j = 1 m log ( P ij ( P ij ) )
Wherein function h 1for representing the computation model of current harmonics feature, c ijit is the current value of the jth subharmonic of i-th electrical equipment.Suppose total n operating electrical equipment, m order harmonics.P ijit is the independent probability distribution of the jth subharmonic of i-th electrical equipment.Total feature value is as above formula.
Consider the current harmonics feature of correlativity between each order harmonics: joint probability distribution when equipment runs between different order harmonic electric current.
The eigenwert of input is the harmonic current of all orders of all electrical equipment;
Output is the logarithm value of a statistical probability, the probability that the eigenwert just inputted therewith for all electrical equipment is identical.
Formula is:
h 2 ( { c ij } ) = log ( Π i = 1 n P i { c ij | j = 1,2 . . . m } ) = Σ i = 1 n log ( P i { c ij | j = 1,2 . . . m } )
Wherein function h 2for representing the computation model of current harmonics correlative character, c ijit is the current value of the jth subharmonic of i-th electrical equipment.Suppose total n operating electrical equipment, m order harmonics.P ibe i-th electrical equipment current harmonics on correlation probabilities distribution.Total feature value is as above formula.
Conductance feature: the conductance probability distribution of equipment when using equipment to run, basic definition is the same, and electric current is changed to conductance
Conductance is defined as follows:
G = R R 2 + X 2 = Re ( i ( t ) v ( t ) )
The real part of admittance when the value of conductance is equipment work, the calculating of getting real part represents with function Re.With two kinds of characterizing definitions are similar above, conductance is characterized as the probability distribution of its conductance when equipment runs, and wherein v (t) and i (t) represent electric current and the voltage of t respectively.
Input feature vector is the conductance of all devices;
Output is the logarithm value of a statistical probability, the probability that the conductance just inputted therewith for all electrical equipment is identical.
h 3 ( { G i } ) = log ( Π i = 1 n P ′ i { G i } ) = Σ i = 1 n log ( P ′ i { G i } )
Wherein, function h 3for representing the computation model of conductance feature, G iit is the electric conductivity value (being calculated by current/voltage) of i-th electrical equipment.P ' iit is the conductance distribution of i-th electrical equipment.
To each electrical equipment, set up the feature corresponding to each statistical model.
After determining the statistical nature needing to measure, the calculating each equipment being carried out to statistical distribution can be started.
Installation testing environment, obtains raw data.
Fig. 6 is statistical nature environment schematic to be measured, according to Fig. 6, in data acquisition system (DAS), installs suitable current/voltage sensor between Devices to test and civil power, builds experimental situation: the data of the collection of processing unit processes data acquisition system (DAS).Afterwards, the voltage when electricity measurer normally works and current data is recorded.The data now gathered are ifq circuit data, can represent electric current and the voltage of t with v (t) and i (t) respectively.
1) according to definition, counting statistics distributes.
According to the definition of three features, first original circuit data is converted to harmonic wave and represents (frequency domain) by us
To moment T, with and subsequent the data of one-period carry out Fourier transform, the harmonic wave that can obtain electric current in this cycle that the T moment starts and voltage represents have
I k ( T ) = Σ n = 0 N - 1 i ( T + n N × 0.02 ) e - j 2 πk n N k = 1,2 . . . N 2
Wherein, I k(T) being the value of harmonic current on kth rank, is a complex value.N is the sampling number in one-period, and n represents the number of electrical equipment.
The computing method of voltage are identical with electric current.
Conductance G = Re ( i ( T ) v ( T ) )
Each cycle of the raw data collected so is calculated, just can obtain a large amount of harmonic current data.
According to characterizing definition, calculate the independent distribution probability P of harmonic current respectively ij, joint distribution probability P i, and the distribution probability P ' of conductance i.
Arrive this, even if the statistical model of equipment has been set up
Two. in the family of user, adjust weight parameter
When weight parameter regulates the stage to start, first need the installment work of analytic system in subscriber household.Fig. 7 is the electricity consumption behavior sensory perceptual system schematic diagram of household internal power utilization environment.System installs collecting unit according to Fig. 7 in the electrical network porch of subscriber household, and collecting unit is from electrical network Acquisition Circuit data (power information), and the result that computing unit obtains according to described circuit data compute dependent data is sent to display unit.
Simply introduce weight parameter below to the impact analyzed:
Use log-linear model in order to multiple feature is combined us effectively, formula is as follows:
P ( c | t ) = exp ( Σ i = 1 m λ i h i ( c | t ) ) Σ c ′ exp ( Σ i = 1 m λ i h i ( c ′ | t ) )
Wherein: h ithe statistical nature of definition for it, and λ ifor character pair h iweight parameter.T and c represents the circuit information recorded, the circuit state analyzed by analytic system respectively.Different physical significances (meaning of t only just can show when the feature of existence condition probability) is brought in different features.Such as: to current harmonics feature, the harmonic current of each order of all devices under this circuit state of c, t is nonsensical.In above formula, the effect of denominator is normalization, and c ' wherein represents all possible situation, when we will obtain most possible c, when P (c|t) is maximized, can remove denominator, have:
c = arg max c ( P ( c | t ) ) = arg max c ( exp ( Σ i = 1 m λ i h i ( c | t ) ) Σ c ′ exp ( Σ i = 1 m λ i h i ( c ′ | t ) ) ) = arg max c ( Σ i = 1 m λ i h i ( c | t ) )
Therefore only obtaining and make each characteristic line and maximum state c, is exactly the state that system is most possibly in.
And weight λ has just indicated the significance level between feature.
Such as: have two different system state c1 and c2; Its characteristic of correspondence value is as following table:
c1 c2
h1 1 3
h2 3 1
h3 1 2
As: λ gets 1 entirely, then c2 is more excellent; λ is that (1,0.5,0.5) then c1 is more excellent.Different parameters can obtain different optimal situation
The adjustment weight parameter stage task that will complete is exactly, and by parameter adjustment to a suitable state, makes the optimal situation calculated by log-linear model, identical as much as possible with the electricity consumption situation of reality.
After environmental structure completes, formally start the parameter learning stage.
1. arrange initial weight, initial weight all can get 1.
2. the circuit data of system acquisition long period, comprises the data in multiple moment.
3., by the method (see Part IV) of search, for each moment calculates a list, deposit the candidate state making P (c|t) large as far as possible.
4. give user by these lists, from each list, select the candidate state the most similar to actual electricity consumption situation by user.
5. start iteration weight optimization algorithm, algorithm flow is as follows:
Input: a series of candidate state list, the optimum state that each list is corresponding, the value of each feature in list under each state, initial weight
Export: the weight after optimization:
Algorithm:
1) choose each feature successively, perform 2) step.
2) adjust this feature respective weights, make in all candidate lists, the optimum state calculated and user input identical as far as possible many.
Complete one take turns optimization after, according to new weight calculation candidate list.As item optimum in each list inputs identical with user, or without any optimization, then exit optimizing process.Otherwise, get back to 1) and step proceeds.
With an example, this process is described below:
Suppose, have three moment, three features, in the candidate list in each moment, have three states.The value of three feature h1, h2, h3 is as shown in the table (shown in table is one group of value supposed):
User inputs optimum state and is respectively: candidate 1, and candidate 5, and candidate 7
Algorithm is by, λ 123=1 starts
First round iteration:
Selected characteristic 1, brings in summation formula, proper λ 1during >1, the moment 1 and 3 can get optimum, and we can suppose λ 1=2.
Selected characteristic 2, brings in summation formula, can obtain 2> λ 2during >1.5, all moment can obtain optimum.
We can suppose λ 2=1.75;
All meet to this all moment.Optimizing process can stop.Final weight is respectively
λ 1=2;λ 2=1.75;λ 3=1
6. export optimized parameter, the parameter learning stage terminates.
Three. the electrical energy consumption analysis stage
After completing model foundation and parameter training, system starts electrical energy consumption analysis.The hardware installation method in this stage is identical with the upper parameter training stage, installs collecting unit in electrical network porch.
In this stage, the circuit data of analytic system to each moment is analyzed.Calculate electricity consumption situation optimum in family respectively, computing method are below in square frame, and the method can also be used for calculating optimum candidate list.
Input: the characteristic model of each electrical equipment, the weight parameter of each model, current and voltage data total in a certain moment family.
Export: (or multiple) optimum state for all electrical equipment in this moment whole family front yard.
Algorithm flow:
1. by total current and voltage data according to characterizing definition, be converted to frequency domain, represent with harmonic wave form.
2. set up three empty queues Closed, Tmp and Open, be respectively used to record possible state, the state that buffer memory is to be deployed, record the state launched.Set up a Hash table H, record the state launched
3. choose an original state S 0, the strategy chosen has two kinds usually, retains the optimum state in a upper moment, or chooses full electrical equipment closed condition, calculates S 0probability and by S 0put into Tmp and H;
4. all elements in queue Tmp is imported Open, empty Tmp;
5. from Open, choose a state S i, calculate all states (adjacent definition is the duty difference only having an equipment) be adjacent, by the part do not appeared in these states in Hash table H, join in Tmp and H.By S imove to Closed from Open
6. repeat step 5, until Open is empty.
7. if Tmp is empty, then stop search, enter step 9.Otherwise to all states in Tmp, calculate the value of each feature under each state, obtain final probability by log-linear model.According to probability size, these states are sorted, only retain top n state (N is the search width preset).
If 8. in Tmp table, all shape probability of states are all less than the optimal value in Closed.Then stop search, enter step 9.Otherwise get back to step 4, continue search.
States all in Tmp moved in Closed, (multiple) state that select probability is maximum from Closed is as the output of algorithm.
(a kind of statistical nature is only used in this example: the current characteristic with correlativity below with a simplified example, a current harmonics tri-vector represents, three dimensions is expressed as a coordinate system of xyz, in this example, electrical appliance state only has two kinds, switch, F represents pass, T represents out), compare analyzing process is demonstrated;
Assuming that there are now three electrical equipment, be A, B, C respectively.Feature on it is respectively:
The distribution of current of electrical equipment A is being uniformly distributed in x-axis, and be a class normal distribution in yz plane, its probability function can be expressed as, and a is constant:
P A ( x , y , z ) = ae - 1 2 ( y 2 + z 2 )
The distribution of current of electrical equipment B is being uniformly distributed in y-axis, and be a class normal distribution in xz plane, its probability function can be expressed as, and b is constant:
P B ( x , y , z ) = be - 1 2 ( x 2 + z 2 )
The distribution of current of electrical equipment C is being uniformly distributed in z-axis, and be a class normal distribution in xy plane, its probability function can be expressed as, and c is constant:
P C ( x , y , z ) = ce - 1 2 ( x 2 + y 2 )
Therefore the account form of feature is:
h ( { c i } ) = Σ i = 1 n log ( P i ( c i ) )
C iit is the current vector of i-th electrical equipment.Only has a feature, so log-linear model just equals the exponential quantity of this feature.
After simplifying, maximized function to be expressed as
f ( { c 1 , c 2 , c 3 } ) = - ( y 1 2 + z 1 2 + x 2 2 + z 2 2 + x 3 2 + z 3 2 )
Start search procedure below, the three-dimensional boolean vector of a state in process represents, as (FTF).Suppose that the width searched for is 2
1. calculate total current harmonic wave to represent, assuming that total current vector is (1,2,3)
2. initialize queue Open, Closed and Tmp, initialization Hash H
3. original state chooses the state S that whole electrical equipment all cuts out 0(FFF), obviously when total current is not 0, S 0probability is 0, and objective function is equivalent to; By S 0add in Tmp and H
4. by the whole states in Tmp, namely S 0join in Open;
5. choose a state (FFF) in Open, adjacent states have (TFF), (FTF), (FFT).Put into Tmp and H, (FFF) puts into Closed table
6. all state computation objective functions in couple Tmp, and retain the first two state
(TFF) under state, obvious all electric currents are all that electrical equipment A produces.So target function value is-13;
(FTF) under state, obvious all electric currents are all that electrical equipment B produces.So target function value is-10;
(FFT) under state, obvious all electric currents are all that electrical equipment A produces.So target function value is-5;
Finally retain (FTF), (FFT),
7. continue search, states all in Tmp are added in Open
8. choose a state (FTF) in Open, adjacent simultaneously not appearing at has (TTF), (FTT) in H.Put into Tmp and H, (FTF) puts into Closed table
9. choose a state (FFT) in Open, adjacently do not appear at having (TFT) in H simultaneously.Put into Tmp and H, (FFT) puts into Closed table
10. all state computation objective functions in couple Tmp, and retain the first two state
(TTF) under state, electric current is all produced by electrical equipment A and electrical equipment B, is respectively (1,0,1.5), maximum probability in time (0,2,1.5) according to characteristic formula A and the B stream that powers on that is easy to get.Target function value is-4.5;
In like manner can obtain, under (FTT) state, B and the C stream that powers on is respectively (0.5,2,0), maximum probability in time (0.5,0,3).Target function value is-0.5;
In like manner can obtain, under (TFT) state, A and the C stream that powers on is respectively (1,1,0), maximum probability in time (0,1,3).Target function value is-1;
Finally retain (FTT), (TFT),
11. continue search, states all in Tmp are added in Open
12. choose a state (FTT) in Open, adjacently do not appear at having (TTT) in H simultaneously.Put into Tmp and H, (FTT) puts into Closed table
13. choose a state (TFT) in Open, there is not the adjacent state simultaneously do not appeared in H, and (TFT) puts into Closed table
All state computation objective functions in 14. couples of Tmp, and retain the first two state
(TTT) under state, electric current is produced by electrical equipment A, electrical equipment B and electrical equipment C, is respectively (1,0,0), (0,2,0), maximum probability in time (0,0,3) according to characteristic formula A, B and the C stream that powers on that is easy to get.Target function value is 0;
15. continue search, states all in Tmp are added in Open
16. choose a state (TTT) in Open, there is not the adjacent state simultaneously do not appeared in H, and (TTT) puts into Closed table
17.Tmp table is for empty, and search stops, and to state sequences all in Closed table, obtaining optimum state is (TTT), and suboptimum is (FTT)
Obviously this result and our understanding are also identical, and the current vector of A, B, C tri-electrical equipment is respectively in three dimensions, and naturally when total current is (1,2,3), three electrical equipment all should be in open mode.
Certainly; the present invention also can have other various embodiments; when not deviating from the present invention's spirit and essence thereof; those of ordinary skill in the art are when making various corresponding change and distortion according to the present invention, but these change accordingly and are out of shape the protection domain that all should belong to the claim appended by the present invention.

Claims (8)

1. a cognitive method for various dimensions electricity consumption behavior, is characterized in that, comprising:
Step 1, is previously defined in the statistical nature of consumer under multiple dimension and the eigenwert under obtaining often kind of statistical nature, and sets up corresponding statistical model;
Step 2, install collecting unit in electrical network porch, the current-voltage information of all consumers of Real-time Obtaining, computing unit is according to the input data of described current-voltage information and user, train the feature weight of each statistical nature, wherein said step 2 comprises:
Step 21, measures the total current information of voltage of all consumers in the family in a period of time, starts to carry out weight training as training dataset;
Step 22, is calculated by heuristic search, obtains the candidate list of multiple candidate result composition;
Step 23, user chooses optimum analysis result according to truth in candidate list;
Step 24, carries out iterative optimization weight parameter by the information provided described candidate list and user, and obtain the feature weight of described all consumers, wherein iterative algorithm comprises step 241, chooses each statistical nature successively, performs step 242; Step 242, adjustment statistical nature respective weights, makes in all candidate lists, the optimum state calculated and user input identical try one's best many, complete one take turns optimization after, according to new weight calculation candidate list;
Step 3, described computing unit carries out heuristic search calculating according to described eigenwert, statistical model and feature weight, obtains the real-time electricity consumption behavior of each consumer, and described real-time electricity consumption behavior is sent to display unit.
2. the cognitive method of various dimensions electricity consumption behavior as claimed in claim 1, it is characterized in that, described collecting unit comprises a current sensor and a voltage sensor, this current sensor is the magnetic field induction chip based on Hall effect, electromagnetic induction is utilized to realize the current data collection of non-intrusion type, the direct place in circuit of this voltage sensor is in parallel with all consumers, measures the voltage on it.
3. the cognitive method of various dimensions electricity consumption behavior as claimed in claim 1, it is characterized in that, described computing unit independently uses a processor; Or share a processor with collecting unit; Or share a processor with display unit.
4. the cognitive method of various dimensions electricity consumption behavior as claimed in claim 1, it is characterized in that, the heuristic computing formula in described step 3 is:
P ( c | t ) = exp Σ i = 1 n λ i h i ( c , t ) Σ c ′ exp Σ i = 1 n λ i h i ( c ′ , t )
Wherein, t represents character representation total in family, and c is the one electricity condition of all electrical equipment in family, and P (c|t) represents when known t, and in family, electricity consumption situation is just the probability of c; The h on right side ibe the fundamental function of a dimension, λ ifor the h of correspondence iweight parameter, denominator is by the summation of the probability of possible state c ' all in the family when t, and n is the inner primary electricity using device number of types of load.
5. a sensory perceptual system for various dimensions electricity consumption behavior, is characterized in that, comprising:
Pretreatment module, is previously defined in the statistical nature of consumer under multiple dimension and the eigenwert under obtaining often kind of statistical nature, and sets up corresponding statistical model;
Training module, in electrical network porch, collecting unit is installed, the current-voltage information of all consumers of Real-time Obtaining, computing unit trains the feature weight of each statistical nature according to the input Data Comparison of described current-voltage information and user, and wherein said training module comprises:
Weight training preparation module, measures the current-voltage information of each consumer in a period of time, starts to carry out weight training as training dataset;
Candidate collection acquisition module, is calculated by heuristic search, obtains the candidate list of multiple candidate result composition;
Choose module, user chooses optimum analysis result according to truth in candidate list;
Weight obtains module, iterative optimization weight parameter is carried out by the information provided described candidate list and user, obtain the feature weight of described all consumers, wherein iterative algorithm comprises and chooses each statistical nature successively, adjustment statistical nature respective weights, makes in all candidate lists, the optimum state calculated and user input identical try one's best many, complete one take turns optimization after, according to new weight calculation candidate list;
Computing module, described computing unit carries out heuristic search calculating according to described eigenwert, statistical model and feature weight, obtains the real-time electricity consumption behavior of each consumer, and described real-time electricity consumption behavior is sent to display unit.
6. the sensory perceptual system of various dimensions electricity consumption behavior as claimed in claim 5, it is characterized in that, described collecting unit comprises a current sensor and a voltage sensor, this current sensor is the magnetic field induction chip based on Hall effect, electromagnetic induction is utilized to realize the current data collection of non-intrusion type, the direct place in circuit of this voltage sensor is in parallel with all consumers, measures the voltage on it.
7. the sensory perceptual system of various dimensions electricity consumption behavior as claimed in claim 5, it is characterized in that, described computing unit independently uses a processor; Or share a processor with collecting unit; Or share a processor with display unit.
8. the sensory perceptual system of various dimensions electricity consumption behavior as claimed in claim 5, it is characterized in that, the heuristic computing formula in described computing module is:
P ( c | t ) = exp Σ i = 1 n λ i h i ( c , t ) Σ c ′ exp Σ i = 1 n λ i h i ( c ′ , t )
Wherein, t represents character representation total in family, and c is the one electricity condition of all electrical equipment in family, and P (c|t) represents when known t, and in family, electricity consumption situation is just the probability of c; The h on right side ibe the fundamental function of a dimension, λ ifor the h of correspondence iweight parameter, denominator is by the summation of the probability of possible state c ' all in the family when t, and n is the inner primary electricity using device number of types of load.
CN201210555210.8A 2012-12-19 2012-12-19 A kind of cognitive method of various dimensions electricity consumption behavior and system Active CN103020459B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201210555210.8A CN103020459B (en) 2012-12-19 2012-12-19 A kind of cognitive method of various dimensions electricity consumption behavior and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201210555210.8A CN103020459B (en) 2012-12-19 2012-12-19 A kind of cognitive method of various dimensions electricity consumption behavior and system

Publications (2)

Publication Number Publication Date
CN103020459A CN103020459A (en) 2013-04-03
CN103020459B true CN103020459B (en) 2015-12-09

Family

ID=47969057

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201210555210.8A Active CN103020459B (en) 2012-12-19 2012-12-19 A kind of cognitive method of various dimensions electricity consumption behavior and system

Country Status (1)

Country Link
CN (1) CN103020459B (en)

Families Citing this family (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103985063A (en) * 2014-05-16 2014-08-13 苏州鸣伦电子科技有限公司 Imaging feedback method of power demand side consumption state
TWI526852B (en) * 2014-07-29 2016-03-21 國立交通大學 Method for counting number of people based on appliance usages and monitoring system using the same
CN106569026A (en) * 2015-10-09 2017-04-19 腾讯科技(深圳)有限公司 Power consumption device power consumption statistics method and system
CN105652118B (en) * 2015-12-29 2019-02-26 国家电网公司 A kind of power grid power budget monitoring method based on load instantaneous energy feature
TW201820246A (en) * 2016-11-23 2018-06-01 財團法人資訊工業策進會 Method for acquiring load operation probability of electric power consumer and method for acquiring load operation probability of electric power consumer group for acquiring load information and total power consumption information of an electric power consumer or an electric power consumer group
CN106789163B (en) * 2016-11-25 2020-08-18 华电智连信达科技(北京)有限公司 Method, device and system for monitoring power utilization information of network equipment
CN108492134A (en) * 2018-03-07 2018-09-04 国网四川省电力公司 The big data user power utilization behavior analysis system integrated based on multicycle regression tree
CN109299082B (en) * 2018-09-13 2021-09-14 北京中科闻歌科技股份有限公司 Big data analysis method and system
CN109813978B (en) * 2018-12-25 2021-04-20 武汉中原电子信息有限公司 Non-intrusive load type identification method integrating transient characteristic and change characteristic between stable states
CN109782086B (en) * 2018-12-25 2020-12-18 武汉中原电子信息有限公司 Non-intrusive load identification method based on multi-dimensional signal analysis
CN110516743A (en) * 2019-08-28 2019-11-29 珠海格力智能装备有限公司 Recognition methods, device, storage medium and the processor of electrical equipment

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102323469A (en) * 2011-07-27 2012-01-18 四川大学 System for monitoring state of harmonic load

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102323469A (en) * 2011-07-27 2012-01-18 四川大学 System for monitoring state of harmonic load

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
基于稳态谐波分析的非侵入式电力负荷分解法的应用研究;赵春柳;《中国优秀硕士学位论文全文数据库-工程科技II辑》;20111215;第2011年卷(第S2期);C042-535 *
非侵入式电力负荷分解与监测;黎鹏;《中国博士学位论文全文数据库-工程科技II辑》;20101215;第2010年卷(第12期);C042-38 *

Also Published As

Publication number Publication date
CN103020459A (en) 2013-04-03

Similar Documents

Publication Publication Date Title
CN103020459B (en) A kind of cognitive method of various dimensions electricity consumption behavior and system
Zhang et al. Review on the research and practice of deep learning and reinforcement learning in smart grids
CN103245881B (en) Power distribution network fault analyzing method and device based on tidal current distribution characteristics
CN106373025B (en) Power information acquisition system based on outlier detection anti-electricity-theft monitoring method in real time
CN105514984B (en) A kind of non-intrusion type load decomposition device of plug and play
CN104966161A (en) Electric energy quality recording data calculating analysis method based on Gaussian mixture model
CN109145949A (en) Non-intrusive electrical load monitoring and decomposition method and system based on integrated study
CN106203531A (en) A kind of transmission line fault sorting technique based on the sparse autocoder of convolution
CN104239963A (en) Method for finding abnormal electric energy meter based on gray GM (1, 1) model
CN106532719A (en) Non-intrusive identification method of non-variable frequency air conditioner based on second harmonic wave of current and reactive power
CN106537356A (en) Systems and methods for maximizing expected utility of signal injection test patterns in utility grids
CN103018611B (en) Non-invasive load monitoring method and system based on current decomposition
CN109285087A (en) A kind of platform area topology identification method accelerated based on NB-IoT and GPU
CN109784388A (en) Stealing user identification method and device
CN105488335A (en) Lyapunov exponent based power system load prediction method and apparatus
Blanco et al. Stochastic harmonic load model of residential users based on measurements
CN106874676B (en) State evaluation method for electric energy metering device
Cortes-Robles et al. Fast-training feedforward neural network for multi-scale power quality monitoring in power systems with distributed generation sources
CN105406461B (en) A kind of load self-adapting dynamic monitoring method of power distribution network power-off event
CN104573313B (en) The acquisition methods and system of customer response model under Peak-valley TOU power price
CN103258144B (en) Online static load modeling method based on data of fault recorder
Zhang Machine learning and big data analytics for the smart grid
Popeangă Data mining smart energy time series
CN109840691A (en) Non-intrusion type subitem electricity estimation method based on deep neural network
CN108170885A (en) Multi-harmonic Sources recognition methods in a kind of power distribution network

Legal Events

Date Code Title Description
PB01 Publication
C06 Publication
SE01 Entry into force of request for substantive examination
C10 Entry into substantive examination
GR01 Patent grant
C14 Grant of patent or utility model