CN103942453A - Intelligent electricity utilization anomaly detection method for non-technical loss - Google Patents

Abstract

The invention discloses an intelligent electricity utilization anomaly detection method for non-technical loss, and belongs to the technical field of power load analysis. The method includes the steps that (1) original data are preprocessed; (2) feature extraction is conducted on sample data; (3) samples are divided into the initial training samples and the optimization samples; (4) real-time data are sampled, and the sample features are extracted to form a test sample; (5) parameter optimization is conducted through a GA to determine the optical ELM parameter value; (6) anomaly detection is conducted by substitution of the optical ELM parameter value, a training sample and a test sample; (7) if the test time is an integer multiple of 72 hours, classification accuracy and the anomaly error detection rate are counted; if the anomaly error detection rate exceeds the set limit value, the step (8) is executed, and if not, the step (4) is executed; (8) the training sample of a user is updated and the step (5) is executed. The intelligent electricity utilization anomaly detection method for non-technical loss is definite in physical conception, clear in thought, easy and convenient to analyze and calculate, and capable of effectively solving the problem of online detection of non-technical loss of arbitrary electricity utilization loads.

Description

A kind of intelligent power method for detecting abnormality for inartful loss

Technical field

The invention belongs to electric load analysis technical field, relate in particular to a kind of intelligent power method for detecting abnormality for inartful loss.

Background technology

Inartful loss (Nontechnical Loss, NTL) be the concept putting forward with respect to technical loss, be often referred to and be sent to the electric energy that the use of user's side is not still valuated, relevant with a series of duplicity electricity consumption behavior with user's stealing of distribution side.Nowadays, inartful loss has become the key factor that affects Utilities Electric Co.'s income.

For inartful loss, many methods that detect about NTL have abroad been proposed, its detection method is very abundant, comprises statistical method, decision tree, artificial neural network, data mining, Knowledge Discovery and optimal path tree etc., but is all offline inspection.Its historical data is only provided by static historical load curve, if the recent flip-flop consumption habit of user is missed the possibility of surveying as abnormal electricity consumption and greatly promotes, for the limited even user of disappearance of historical data, does not have applicability simultaneously.Along with the development of intelligent power, the universal of intelligent electric meter makes online detection become possibility as the major way of following abnormality detection.Meanwhile, classification accuracy is paid close attention in existing research more, but in the situation that sample radix is very large, the problem that false drop rate is high is not allowed to ignore.On-site verification labor intensive material resources, what flase drop caused is wasted in the situation that sample radix is very large, and the loss bringing is equally very large.Meanwhile, abnormal false drop rate is high, and the algorithm that is suitable for ability of environment is difficult for promoting.

Summary of the invention

The problem existing for above-mentioned prior art, the present invention proposes a kind of intelligent power method for detecting abnormality for inartful loss, it is characterized in that, and the concrete steps of this detection method are:

Step 1: original loads data are carried out to pre-service;

Step 2: sample load data is carried out to feature extraction;

Step 3: sample is divided, determined initial training sample and optimizing sample;

Step 4: utilize Genetic Algorithms to carry out offline parameter optimizing, determine optimum extreme learning machine ELM parameter value;

Step 5: sampling real time data, by the feature of the method extraction sample of step 2, generates test sample book;

Step 6: based on extreme learning machine ELM algorithm, the optimum extreme learning machine ELM of substitution parameter value, initial training sample and test sample book are carried out online abnormality detection;

Step 7: if the detection moment is the integral multiple of 72 hours, statistical classification precision CA and abnormal false drop rate FDC; If exceeding, abnormal false drop rate sets limit value FDC ^limjump to step 8, otherwise, jump to step 5;

Step 8: upgrade user's training sample, jump to step 4.

Described step 1 is specially:

Original loads data are deleted to choosing, delete the original loads data in non-integral point moment, unified samples load data sample frequency;

For original loads shortage of data problem, adopt and fill or cover with the mean value of periodic samples load data, its computing method are:

$x_i=\frac{x_{i-W}+x_{i+W}}{2};W=7\×24;$

Wherein, x _irepresent the sample load data in i moment; W is the theoretical sample number of a week.

Described step 2 is specially:

Primitive character collection to sample carries out pre-service, is converted into output characteristic collection;

Described primitive character collection is by master data element composition, for primitive character is concentrated the sample in m user's i moment, each all a vector that comprises H eigenwert:

$x_i^{m.0}=\{x_h^0;h=1~H\},H=15;$

Primitive character collection comprises sampling time, sampling terminal number, active power, reactive power, A phase current, B phase current, C phase current, A phase voltage, B phase voltage, C phase voltage, 15 minutes average active powers, power factor, accumulative total active power, logical address and family number;

Described output characteristic collection is by master data element composition, for output characteristic is concentrated the sample in m user's i moment, each all a vector that comprises G eigenwert:

$x_i^m=f\left(x_i^{m.0}\right)=\{x_g;g=1~G\};G=16;$

Output characteristic collection comprise active power, reactive power, power factor, accumulative total active power, day peak load, day peak load utilize hour, with electrostrictive coefficient, individual coefficient of dispersion, public coefficient of dispersion, coefficient of dispersion difference, the last hour power consumption ratio same period, same hour of yesterday the power consumption ratio same period, same hour of last week the power consumption ratio same period, peak valley divide, peak valley consecutive numbers and electric weight fluctuation;

The described electrostrictive coefficient of using computing formula be:

${\mathrm{PKAP}}_i^m=\frac{{\mathrm{CAP}}_i^m-{\mathrm{CAP}}_{i-1}^m}{{\mathrm{SEC}}^m};$

Wherein, represent the accumulative total active power of i moment sample of m user; represent the accumulative total active power of i-1 moment sample of m user; SEC ^mrepresent the common transformer capacity of m user place power supply area;

Described individual coefficient of dispersion computing formula be;

$\begin{array}{c}{\mathrm{DC}}_i^m=\sqrt{\frac{1}{n-1}\underset{t=i-n}{\overset{i}{\mathrm{\Σ}}}({\mathrm{AP}}_t^m-\stackrel{\‾}{{\mathrm{AP}}_t^m})}/\stackrel{\‾}{{\mathrm{AP}}_t^m}\\ \stackrel{\‾}{{\mathrm{AP}}_t^m}=\mathrm{avg}({\mathrm{AP}}_{i-n}^m,\·\·\·,{\mathrm{AP}}_i^m)\end{array};;$

Wherein, represent the individual coefficient of dispersion of i moment sample of m user; represent the active power of t moment sample of m user; Avg is the function of calculating mean value, be illustrated in the active power average of analyzing m user in the period; N is 24;

Described public coefficient of dispersion account form be to calculate the average of all users' individual coefficient of dispersion in same power supply area;

Described day peak load computing formula be:

${\mathrm{AP}}_i^m{|}_{\max }=\max ({\mathrm{AP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24}^m,{\mathrm{AP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24+1}^m,...,{\mathrm{AP}}_t^m,...,{\mathrm{AP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24+23}^m);$

Wherein, int is bracket function;

Within described day, peak load is utilized hour t _hourcomputing formula be:

$t_{\mathrm{hour}}=\frac{{\mathrm{CAP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24+23}^m-{\mathrm{CAP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24}^m}{{\mathrm{AP}}_i^m{|}_{\max }};$

Described coefficient of dispersion difference computing formula be:

${\mathrm{\ΔDC}}_i^m=|D{C}_i^m-{\mathrm{PDC}}_i^m|;$

The described power consumption ratio same period computing formula be:

${\mathrm{PC}}_t^m=({\mathrm{AP}}_{t_c}^m-{\mathrm{AP}}_t^m)/{\mathrm{AP}}_{t_c}^m;$

Wherein, represent m user t _cthe active power of individual moment sample; t _crepresent the period being compared the same period, t _ccomprise last hour, same hour of yesterday or same hour of last week;

Described peak valley consecutive numbers QC is the count tag of peak, paddy period length, and peak valley consecutive numbers QC initial value is set to 0; When time, the property value that peak valley is divided be 1, peak valley consecutive numbers QC adds 1; When time, the property value that peak valley is divided be zero; When the property value that peak valley is divided for-1, peak valley consecutive numbers QC subtracts 1; If reindexing, peak valley consecutive numbers QC zero setting again;

Described electric weight fluctuation computing formula be:

${\mathrm{VAR}}_t^m=|{\mathrm{AP}}_t^m-\stackrel{\‾}{{\mathrm{AP}}_t^m}|.$

Described step 3 is specially:

Choose sample after the T moment as preferably foundation of parameter, corresponding optimizing schedule of samples is shown:

$V_T^m=\left\{(l_i,x_i^m)\right\};i=T,T+1,...;$

Wherein, it is the optimizing sample of m user's abnormality detection model; l _ivalue is 1 or-1, represents respectively i the normal electricity consumption of moment sample or abnormal electricity consumption;

Forerunner's sample in corresponding selection T moment is as the initial training sample of abnormality detection model, and it is expressed as:

$M_T^{m.0}=\left\{(l_i,x_i^m)\right\};i=\mathrm{1,2},...,T-1;$

Wherein, it is the initial training sample of m user's abnormality detection model.

Fitness function in described step 4 in the optimizing of Genetic Algorithms offline parameter is taken as abnormality detection accuracy rate ADC, and computing formula is:

$\mathrm{ADC}=d/(b+d+\frac{1}{2});$

Wherein, b represents to be detected as being normally but in fact abnormal sample size by extreme learning machine ELM algorithm; It is to be abnormal sample size extremely and really that d represents to be detected by extreme learning machine ELM algorithm.

Nicety of grading CA in described step 7 is expressed as:

CA＝(a ₁+d ₁)/(a ₁+b ₁+c ₁+d ₁)；

Abnormal false drop rate FDC is expressed as:

FDC＝(c ₁+d ₁+1)/(b ₁+d ₁+1)；

Wherein, on-line stage, a ₁represent to be detected as being normally and really normal sample size by extreme learning machine ELM algorithm; b ₁expression is normally by the detection of extreme learning machine ELM algorithm but is abnormal sample size in fact; c ₁represent to be detected as extremely but be normal sample size in fact by extreme learning machine ELM algorithm; d ₁representing to be detected by extreme learning machine ELM algorithm is to be abnormal sample size extremely and really.

The beneficial effect of the invention: the method for detecting abnormality that the present invention proposes, using abnormal false drop rate as the trigger mechanism that upgrades user's training sample and ELM parameter, detect the index of effect using nicety of grading as assessment, clear physical concept, clear thinking, analytical calculation is easy, can effectively solve the problem of any power load being carried out to the online detection of inartful loss.

Brief description of the drawings

Fig. 1 is the intelligent power abnormality detection schematic diagram for inartful loss;

Fig. 2 is the process flow diagram of the intelligent power method for detecting abnormality that proposes of the present invention.

Embodiment

Below in conjunction with accompanying drawing, elaborate to detecting embodiment.Should be emphasized that, following explanation is only exemplary, instead of in order to limit the scope of the invention and to apply.

This method combines offline inspection and online measuring technique, proposes a kind of intelligent power method for detecting abnormality for inartful loss, and it detects principle as shown in Figure 1.

Fig. 2 is the process flow diagram of the intelligent power method for detecting abnormality that proposes of the present invention, and concrete steps are:

Step 1: original loads data are carried out to pre-service.

Due to different user data acquisition mode disunities, for example commercial user's data be statistics per hour once, industrial user be part statistics per hour once, per half an hour of part or every 15 minutes are added up once, therefore, need to delete choosing to original loads data, delete the original loads data in non-integral point moment, unified samples load data sample frequency is 1 hour/time, is convenient to further analysis.

For original loads shortage of data problem, adopt and fill or cover with the mean value of periodic samples load data, its computing method are:

$x_i=\frac{x_{i-W}+x_{i+W}}{2};W=7\×24;$

Wherein, x _irepresent the sample load data in i moment; W is the theoretical sample number of a week.

Step 2: sample load data is carried out to feature extraction.

Sample load data is carried out to feature extraction, exactly the primitive character collection of sample is carried out to pre-service, be then converted into output characteristic collection.

Primitive character collection is by master data element composition, for primitive character is concentrated the sample in m user's i moment, each all a vector that comprises H eigenwert:

$x_i^{m.0}=\{x_h^0;h=1~H\},H=15;$

Primitive character collection comprises sampling time, sampling terminal number, active power, reactive power, A phase current, B phase current, C phase current, A phase voltage, B phase voltage, C phase voltage, 15 minutes average active powers, power factor, accumulative total active power, logical address and family number.

Output characteristic collection is by master data element composition, for output characteristic is concentrated the sample data in m user's i moment, each all a vector that comprises G eigenwert:

$x_i^m=f\left(x_i^{m.0}\right)=\{x_g;g=1~G\};G=16;$

Output characteristic collection comprise active power, reactive power, power factor, accumulative total active power, day peak load, day peak load utilize hour, with electrostrictive coefficient, individual coefficient of dispersion, public coefficient of dispersion, coefficient of dispersion difference, the last hour power consumption ratio same period, same hour of yesterday the power consumption ratio same period, same hour of last week the power consumption ratio same period, peak valley divide, peak valley consecutive numbers and electric weight fluctuation.

Use electrostrictive coefficient computing formula be:

${\mathrm{PKAP}}_i^m=\frac{{\mathrm{CAP}}_i^m-{\mathrm{CAP}}_{i-1}^m}{{\mathrm{SEC}}^m};$

Wherein, represent the ratio of the common transformer capacity of i moment power consumption of m user and place power supply area; represent the accumulative total active power of i moment sample of m user; represent the accumulative total active power of i-1 moment sample of m user; SEC ^mrepresent the common transformer capacity of m user place power supply area;

Individual's coefficient of dispersion computing formula be;

$\begin{array}{c}{\mathrm{DC}}_i^m=\sqrt{\frac{1}{n-1}\underset{t=i-n}{\overset{i}{\mathrm{\Σ}}}({\mathrm{AP}}_t^m-\stackrel{\‾}{{\mathrm{AP}}_t^m})}/\stackrel{\‾}{{\mathrm{AP}}_t^m}\\ \stackrel{\‾}{{\mathrm{AP}}_t^m}=\mathrm{avg}({\mathrm{AP}}_{i-n}^m,\·\·\·,{\mathrm{AP}}_i^m)\end{array};;$

Wherein, represent the individual coefficient of dispersion of i moment sample of m user; represent the active power of t moment sample of m user; Avg is the function of calculating mean value, be illustrated in the active power average of analyzing user in the period; The data of coefficient of dispersion based on first 24 hours are tried to achieve, and therefore n value is 24.

Public coefficient of dispersion account form be to calculate the average of all users' individual coefficient of dispersion in same power supply area;

Day peak load computing formula be:

${\mathrm{AP}}_i^m{|}_{\max }=\max ({\mathrm{AP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24}^m,{\mathrm{AP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24+1}^m,...,{\mathrm{AP}}_t^m,...,{\mathrm{AP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24+23}^m);$

Wherein, int is bracket function.

Day peak load utilization hour t _hourcomputing formula be:

$t_{\mathrm{hour}}=\frac{{\mathrm{CAP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24+23}^m-{\mathrm{CAP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24}^m}{{\mathrm{AP}}_i^m{|}_{\max }};$

Coefficient of dispersion difference computing formula be:

${\mathrm{\ΔDC}}_i^m=|D{C}_i^m-{\mathrm{PDC}}_i^m|;$

The power consumption ratio same period computing formula be:

${\mathrm{PC}}_t^m=({\mathrm{AP}}_{t_c}^m-{\mathrm{AP}}_t^m)/{\mathrm{AP}}_{t_c}^m;$

Wherein, represent m user t _cthe active power of individual moment sample; t _crepresent the period being compared the same period, t _ccomprise last hour, same hour of yesterday or same hour of last week.

Peak valley consecutive numbers QC is the count tag of peak, paddy period length, and peak valley consecutive numbers QC initial value is set to 0; time, the property value that peak valley is divided be 1, peak valley consecutive numbers QC adds 1; When time, the property value that peak valley is divided be zero; When the property value that peak valley is divided for-1, peak valley consecutive numbers QC subtracts 1; If reindexing, peak valley consecutive numbers QC zero setting again, and so forth.

Electric weight fluctuation computing formula be:

${\mathrm{VAR}}_t^m=|{\mathrm{AP}}_t^m-\stackrel{\‾}{{\mathrm{AP}}_t^m}|.$

Step 3: sample is divided, determined initial training sample and optimizing sample.

Choose sample after the T moment as preferably foundation of parameter, corresponding optimizing schedule of samples is shown:

$V_T^m=\left\{(l_i,x_i^m)\right\};i=T,T+1,...;$

Wherein, it is the optimizing sample of m user's abnormality detection model; l _ivalue is 1 or-1, represents respectively i the normal electricity consumption of moment sample or abnormal electricity consumption;

Forerunner's sample in corresponding selection T moment is as the initial training sample of abnormality detection model, and it is expressed as:

$M_T^{m.0}=\left\{(l_i,x_i^m)\right\};i=\mathrm{1,2},...,T-1;$

Wherein, it is the initial training sample of m user's abnormality detection model.

Step 4: utilize Genetic Algorithms to carry out offline parameter optimizing, determine optimum extreme learning machine ELM parameter value.

Parameter preferred process based on genetic algorithm can be summarized as:

(1) initialization: generate at random one group of initial parameter C ⁰and L ⁰, initial parameter interval is set as respectively [0.1,100] and [0.001,100], and each initial parameter is encoded, and then structure initial population.

(2) fitness assessment: the substitution of extreme learning machine parameter is tried to achieve to corresponding fitness function, and fitness function is taken as abnormality detection accuracy rate ADC, represents that the correct sample number of abnormality detection accounts for the ratio that all detections are exceptional sample.

Try to achieve abnormality detection result according to formula below:

$\mathrm{CL}=\mathrm{ELM}(M_T^{m.0},V_T^m,C^{g_1},L^{g_1}),g_1=\mathrm{1,2,3},...;$

Wherein, the function of classification results, g are obtained in expression based on given initial parameter and optimizing sample ₁represent the training algebraically of genetic algorithm, with represent training g ₁the parameter obtaining after generation, CL (Classification Labels) is corresponding classification results, CL value is 1 or-1, represents respectively to be detected as normal or abnormal by extreme learning machine ELM algorithm.

The truth of sample and the abnormal conditions that detect are made comparisons, draw result as shown in table 1:

The truth of table 1 sample and the abnormal conditions comparative result detecting

Make a represent to be detected as being normally and really normal sample size by ELM algorithm; B represents to be detected as being normally but in fact abnormal sample size by ELM algorithm; C represents to be detected as being extremely but in fact normal sample size by ELM algorithm; It is to be abnormal sample size extremely and really that d represents to be detected by ELM algorithm.The computing formula of fitness function ADC is:

$\mathrm{ADC}=d/(b+d+\frac{1}{2}).$

(3) select, intersect: choose individuality that some fitness are higher as parent population, use heuristic intersection function, move a segment distance as the filial generation intersecting to form, according to scale-up factor to poor parent

$P_1^{g_1+1}=P_1^{g_1}+R(P_2^{g_1}-P_1^{g_1});$

Wherein, R is scale-up factor, and value is the random number in (1,2) scope, for parent parameter value, wherein for the higher parent of fitness, for filial generation parameter value.

$P_1^{g_1+1}=P_1^{g_1+1}+\mathrm{rand}\left(g\right);$

Wherein, rand function is the random function that produces random number.

(4) detect: change while being less than 1% when population algebraically exceedes maximum algebraically or the fitness in continuous 10 generations, terminal parameter searching process, exports current optimal value of the parameter C ^optand L ^opt; Otherwise jump procedure (2).

Step 5: sampling real time data, by the feature of the method extraction sample of step 2, generates test sample book.

The frequency of on-line sampling is once per hour, in the time carrying out online abnormality detection, detects moment T ₀primitive character sample be:

${\mathrm{TE}}_t^{m.0}=\{x_h^0;h=1~H\};t=T_0,H=15;$

Wherein, it is the primitive character sample in m user's t moment.

Consistent with the disposal route of raw data in step 2, the T of on-line sampling ₀the test sample book in moment will be converted into primitive character collection the output characteristic collection of appointment equally,

${\mathrm{TE}}_t^m=f\left({\mathrm{TE}}_t^{m.0}\right)=\{x_g;g=1~G\};t=T_0,G=16;$

Wherein, it is the output characteristic sample in m user's t moment.

Step 6: based on extreme learning machine ELM algorithm, substitution extreme learning machine ELM optimal value of the parameter, training sample and test sample book are carried out online abnormality detection.

Based on extreme learning machine ELM algorithm, operating limit learning machine ELM optimal value of the parameter, training sample and test sample book are set up abnormality detection model, and formula is:

${\hat{l}}_t=\mathrm{ELM}(M^{m.k},{\mathrm{TE}}_t^m,C^{m.k},L^{m.k});k=\mathrm{1,2,3}...;$

Wherein, be the testing result in t moment, its value is 1 or-1; K is online updating number of times; M ^m.kit is the training sample of the abnormality detection model after m user upgrades for the k time; for the output characteristic of real time data is concentrated the test sample book in t moment of m user; C ^m.k, L ^m.kfor Genetic Algorithms try to achieve corresponding to training sample M ^m.koptimal value of the parameter.

Step 7: if the detection moment is the integral multiple of 72 hours, statistical classification precision and abnormal false drop rate index; If suddenly unprecedented soaring exceeding of abnormal false drop rate FDC set limit value FDC ^lim, illustrate that user has found great accident, as production and operation direction adjustment etc., now jump to step 8; Otherwise, jump to step 4.

Nicety of grading CA has been widely used in correlative study as traditional evaluation criterion, represents Detection accuracy, and expression is:

CA＝(a ₁+d ₁)/(a ₁+b ₁+c ₁+d ₁)；

Wherein, on-line stage, a ₁represent to be detected as being normally and really normal sample size by extreme learning machine ELM algorithm; b ₁expression is normally by the detection of extreme learning machine ELM algorithm but is abnormal sample size in fact; c ₁represent to be detected as extremely but be normal sample size in fact by extreme learning machine ELM algorithm; d ₁representing to be detected by extreme learning machine ELM algorithm is to be abnormal sample size extremely and really.

After abnormal electricity consumption behavior being detected, being detected as abnormal electricity consumption data conventionally need to carry out artificial on-the-spot confirmation and process, and site inspection need to consume a large amount of manpowers and material resources, therefore select abnormal false drop rate FDC another evaluation index as algorithm performance.The sample proportion that this index is abnormal user by statistics by false retrieval, reflects financial cost and human resources value that Outlier Detection Algorithm brings, can be expressed as:

FDC＝(c ₁+d ₁+1)/(b ₁+d ₁+1)。

Step 8: upgrade user's training sample, jump to step 5.

When the suddenly unprecedented soaring setting value of crossing of abnormal false drop rate FDC, illustrate that great accident has occurred user, as management direction adjustment etc.Now, based on the test sample book online updating training sample after history samples and checking, the test sample book of the most recent 72 hours is updated to training sample, delete in former training sample the sample data of the most first 72 hours simultaneously, be stabilized in a fixing numerical value to maintain number of training, concrete formula is:

M ^m.k＝M ^m.(k-1)+TE ^m.k-M _k

${\mathrm{TE}}^{m.k}=\left\{(l_{t_i},{\mathrm{TM}}_{t_i}^m)\right\},t_i=t-71,t-70,...,t;$

$M_k=\left\{(l_i,x_i^m)\right\}\⋐M_T^m;i=72k-71,\·\·\·,72k-\mathrm{1,72}k$

Wherein, M ^{m. (k-1)}it is the training sample of the abnormality detection model after m user upgrades for the k-1 time; TE ^m.krepresent that m user upgrades for the k time in the process of training sample, the set of the test sample book of youngest 72 hours of extracting; represent m user t _ithe test sample book in moment; M _kthe sample data collection of the most first 72 hours in the former training sample of indicating to delete, it is the training sample M of gained after upgrading for k-1 time ^{m. (k-1)}subset.

The present invention goes for the on-line monitoring of various types of customer charge data, comprises and is not limited to the electricity consumption data of the types such as historical data shortage, user's flip-flop consumption habit.

The above; only for preferably embodiment of the present invention, but protection scope of the present invention is not limited to this, is anyly familiar with in technical scope that those skilled in the art disclose in the present invention; the variation that can expect easily or replacement, within all should being encompassed in protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of claim.

Claims (6)

1. for an intelligent power method for detecting abnormality for inartful loss, it is characterized in that, the concrete steps of this detection method are:

Step 1: original loads data are carried out to pre-service;

Step 2: sample load data is carried out to feature extraction;

Step 3: sample is divided, determined initial training sample and optimizing sample;

Step 4: utilize Genetic Algorithms to carry out offline parameter optimizing, determine optimum extreme learning machine ELM parameter value;

Step 5: sampling real time data, by the feature of the method extraction sample of step 2, generates test sample book;

Step 6: based on extreme learning machine ELM algorithm, the optimum extreme learning machine ELM of substitution parameter value, initial training sample and test sample book are carried out online abnormality detection;

Step 7: if the detection moment is the integral multiple of 72 hours, statistical classification precision CA and abnormal false drop rate FDC; If exceeding, abnormal false drop rate sets limit value FDC ^limjump to step 8, otherwise, jump to step 5;

Step 8: upgrade user's training sample, jump to step 4.

2. a kind of intelligent power method for detecting abnormality for inartful loss according to claim 1, is characterized in that, described step 1 is specially:

Original loads data are deleted to choosing, delete the original loads data in non-integral point moment, unified samples load data sample frequency;

For original loads shortage of data problem, adopt and fill or cover with the mean value of periodic samples load data, its computing method are:

$x_i=\frac{x_{i-W}+x_{i+W}}{2};W=7\×24;$

Wherein, x _irepresent the sample load data in i moment; W is the theoretical sample number of a week.

3. a kind of intelligent power method for detecting abnormality for inartful loss according to claim 2, is characterized in that, described step 2 is specially:

Primitive character collection to sample carries out pre-service, is converted into output characteristic collection;

Described primitive character collection is by master data element composition, for primitive character is concentrated the sample in m user's i moment, each all a vector that comprises H eigenwert:

$x_i^{m.0}=\{x_h^0;h=1~H\},H=15;$

Primitive character collection comprises sampling time, sampling terminal number, active power, reactive power, A phase current, B phase current, C phase current, A phase voltage, B phase voltage, C phase voltage, 15 minutes average active powers, power factor, accumulative total active power, logical address and family number;

Described output characteristic collection is by master data element composition, for output characteristic is concentrated the sample in m user's i moment, each all a vector that comprises G eigenwert:

$x_i^m=f\left(x_i^{m.0}\right)=\{x_g;g=1~G\};G=16;$

Output characteristic collection comprise active power, reactive power, power factor, accumulative total active power, day peak load, day peak load utilize hour, with electrostrictive coefficient, individual coefficient of dispersion, public coefficient of dispersion, coefficient of dispersion difference, the last hour power consumption ratio same period, same hour of yesterday the power consumption ratio same period, same hour of last week the power consumption ratio same period, peak valley divide, peak valley consecutive numbers and electric weight fluctuation;

The described electrostrictive coefficient of using computing formula be:

${\mathrm{PKAP}}_i^m=\frac{{\mathrm{CAP}}_i^m-{\mathrm{CAP}}_{i-1}^m}{{\mathrm{SEC}}^m};$

Wherein, represent the accumulative total active power of i moment sample of m user; represent the accumulative total active power of i-1 moment sample of m user; SEC ^mrepresent the common transformer capacity of m user place power supply area;

Described individual coefficient of dispersion computing formula be;

$\begin{array}{c}{\mathrm{DC}}_i^m=\sqrt{\frac{1}{n-1}\underset{t=i-n}{\overset{i}{\mathrm{\Σ}}}({\mathrm{AP}}_t^m-\stackrel{\‾}{{\mathrm{AP}}_t^m})}/\stackrel{\‾}{{\mathrm{AP}}_t^m}\\ \stackrel{\‾}{{\mathrm{AP}}_t^m}=\mathrm{avg}({\mathrm{AP}}_{i-n}^m,\·\·\·,{\mathrm{AP}}_i^m)\end{array};;$

Wherein, represent the individual coefficient of dispersion of i moment sample of m user; represent the active power of t moment sample of m user; Avg is the function of calculating mean value, be illustrated in the active power average of analyzing m user in the period; N is 24;

Described public coefficient of dispersion account form be to calculate the average of all users' individual coefficient of dispersion in same power supply area;

Described day peak load computing formula be:

${\mathrm{AP}}_i^m{|}_{\max }=\max ({\mathrm{AP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24}^m,{\mathrm{AP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24+1}^m,...,{\mathrm{AP}}_t^m,...,{\mathrm{AP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24+23}^m);$

Wherein, int is bracket function;

Within described day, peak load is utilized hour t _hourcomputing formula be:

$t_{\mathrm{hour}}=\frac{{\mathrm{CAP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24+23}^m-{\mathrm{CAP}}_{\mathrm{int}\left(\frac{t}{24}\right)\·24}^m}{{\mathrm{AP}}_i^m{|}_{\max }};$

Described coefficient of dispersion difference computing formula be:

${\mathrm{\ΔDC}}_i^m=|D{C}_i^m-{\mathrm{PDC}}_i^m|;$

The described power consumption ratio same period computing formula be:

${\mathrm{PC}}_t^m=({\mathrm{AP}}_{t_c}^m-{\mathrm{AP}}_t^m)/{\mathrm{AP}}_{t_c}^m;$

Wherein, represent m user t _cthe active power of individual moment sample; t _crepresent the period being compared the same period, t _ccomprise last hour, same hour of yesterday or same hour of last week;

Described peak valley consecutive numbers QC is the count tag of peak, paddy period length, and peak valley consecutive numbers QC initial value is set to 0; When time, the property value that peak valley is divided be 1, peak valley consecutive numbers QC adds 1; When time, the property value that peak valley is divided be zero; When the property value that peak valley is divided for-1, peak valley consecutive numbers QC subtracts 1; If reindexing, peak valley consecutive numbers QC zero setting again;

Described electric weight fluctuation computing formula be:

${\mathrm{VAR}}_t^m=|{\mathrm{AP}}_t^m-\stackrel{\‾}{{\mathrm{AP}}_t^m}|.$

4. a kind of intelligent power method for detecting abnormality for inartful loss according to claim 3, is characterized in that, described step 3 is specially:

Choose sample after the T moment as preferably foundation of parameter, corresponding optimizing schedule of samples is shown:

$V_T^m=\left\{(l_i,x_i^m)\right\};i=T,T+1,...;$

Wherein, it is the optimizing sample of m user's abnormality detection model; l _ivalue is 1 or-1, represents respectively i the normal electricity consumption of moment sample or abnormal electricity consumption;

Forerunner's sample in corresponding selection T moment is as the initial training sample of abnormality detection model, and it is expressed as:

$M_T^{m.0}=\left\{(l_i,x_i^m)\right\};i=\mathrm{1,2},...,T-1;$

Wherein, it is the initial training sample of m user's abnormality detection model.

5. a kind of intelligent power method for detecting abnormality for inartful loss according to claim 4, is characterized in that, the fitness function in described step 4 in the optimizing of Genetic Algorithms offline parameter is taken as abnormality detection accuracy rate ADC, and computing formula is:

$\mathrm{ADC}=d/(b+d+\frac{1}{2});$

Wherein, b represents to be detected as being normally but in fact abnormal sample size by extreme learning machine ELM algorithm; It is to be abnormal sample size extremely and really that d represents to be detected by extreme learning machine ELM algorithm.

6. a kind of intelligent power method for detecting abnormality for inartful loss according to claim 5, is characterized in that, the nicety of grading CA in described step 7 is expressed as:

CA＝(a ₁+d ₁)/(a ₁+b ₁+c ₁+d ₁)；

Abnormal false drop rate FDC is expressed as:

FDC＝(c ₁+d ₁+1)/(b ₁+d ₁+1)；

Wherein, on-line stage, a ₁represent to be detected as being normally and really normal sample size by extreme learning machine ELM algorithm; b ₁expression is normally by the detection of extreme learning machine ELM algorithm but is abnormal sample size in fact; c ₁represent to be detected as extremely but be normal sample size in fact by extreme learning machine ELM algorithm; d ₁representing to be detected by extreme learning machine ELM algorithm is to be abnormal sample size extremely and really.