CN109165604A - The recognition methods of non-intrusion type load and its test macro based on coorinated training - Google Patents
The recognition methods of non-intrusion type load and its test macro based on coorinated training Download PDFInfo
- Publication number
- CN109165604A CN109165604A CN201810989105.2A CN201810989105A CN109165604A CN 109165604 A CN109165604 A CN 109165604A CN 201810989105 A CN201810989105 A CN 201810989105A CN 109165604 A CN109165604 A CN 109165604A
- Authority
- CN
- China
- Prior art keywords
- load
- wavelet
- current signal
- transient current
- identification method
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 55
- 238000012549 training Methods 0.000 title claims abstract description 40
- 238000012360 testing method Methods 0.000 title claims abstract description 20
- 238000003066 decision tree Methods 0.000 claims abstract description 35
- 230000001052 transient effect Effects 0.000 claims abstract description 34
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 22
- 238000013461 design Methods 0.000 claims abstract description 16
- 238000003064 k means clustering Methods 0.000 claims description 4
- 238000001914 filtration Methods 0.000 claims description 3
- 230000008901 benefit Effects 0.000 abstract description 9
- 238000004364 calculation method Methods 0.000 abstract description 3
- 230000009286 beneficial effect Effects 0.000 abstract 1
- 238000007418 data mining Methods 0.000 abstract 1
- 230000005611 electricity Effects 0.000 description 14
- 230000009466 transformation Effects 0.000 description 7
- 230000006870 function Effects 0.000 description 5
- 238000004458 analytical method Methods 0.000 description 4
- 238000013480 data collection Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000002592 echocardiography Methods 0.000 description 3
- 101100499229 Mus musculus Dhrsx gene Proteins 0.000 description 2
- 238000012512 characterization method Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 235000013399 edible fruits Nutrition 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 238000000342 Monte Carlo simulation Methods 0.000 description 1
- 235000006508 Nelumbo nucifera Nutrition 0.000 description 1
- 240000002853 Nelumbo nucifera Species 0.000 description 1
- 235000006510 Nelumbo pentapetala Nutrition 0.000 description 1
- HUTDUHSNJYTCAR-UHFFFAOYSA-N ancymidol Chemical compound C1=CC(OC)=CC=C1C(O)(C=1C=NC=NC=1)C1CC1 HUTDUHSNJYTCAR-UHFFFAOYSA-N 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000005520 cutting process Methods 0.000 description 1
- 238000013499 data model Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 239000000686 essence Substances 0.000 description 1
- 239000004615 ingredient Substances 0.000 description 1
- 230000009191 jumping Effects 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000011426 transformation method Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Signal Processing (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
本发明公开了一种非侵入式负荷识别方法及其测试系统，基于小波设计和数据挖掘协同训练，通过一对余算法，结合决策树模型和k最近邻分类器对瞬态电流信号数据集进行协同训练，以对数据集中的各瞬态电流信号进行分类，并据此预测各负载的类别，从而实现了以下有益效果：1.小波变换能较好的定位待识别信号的波形特征，具有多分辨率的特点，在时域和频域都具有表征信号局部特征的能力，利于提取负载有效的特征信息。2.基于k最近邻分类器和决策树模型协同训练的算法不但降低了计算的复杂度，而且提高了预测精度。3.随着负载数量的增加，该识别方法具有计算时间短，可靠性高的优势。
The invention discloses a non-intrusive load identification method and a testing system thereof. Based on wavelet design and data mining collaborative training, the transient current signal data set is subjected to a pair of redundant algorithm combined with a decision tree model and a k-nearest neighbor classifier. Co-training is used to classify each transient current signal in the data set and predict the type of each load accordingly, thereby achieving the following beneficial effects: 1. Wavelet transform can better locate the waveform characteristics of the signal to be identified, and has many The characteristic of resolution is that it has the ability to characterize the local characteristics of the signal in both the time domain and the frequency domain, which is conducive to extracting the effective feature information of the load. 2. The algorithm based on the co-training of the k-nearest neighbor classifier and the decision tree model not only reduces the computational complexity, but also improves the prediction accuracy. 3. As the number of loads increases, the identification method has the advantages of short calculation time and high reliability.
Description
Technical field
The invention belongs to smart grid and field of signal processing more particularly to a kind of non-intrusion type based on coorinated training are negative
Lotus recognition methods and its test macro.
Background technique
Non-intrusion type load identification technology is a kind of by obtaining the different essences of user at outdoor panel or intelligent electric meter
The power information of thin degree, is an important research direction of intelligent power field, by obtaining the changing rule of load ingredient,
Accurate time-varying load model could be established, to grid company realize fining demand side management, accurate load prediction and
The dynamic characteristic of analysis active distribution system is of great significance.
The key for solving non-intrusion type load identification technology is to analyze the transient state or stable state extracted in electric appliance use process
Signal.Extensive and in-depth research and discussion have been unfolded to the extracting method of signal both at home and abroad.Such as short time discrete Fourier transform, S becomes
It changes, the technologies such as wavelet transformation all have respective deficiency: although (1) Fourier transform has the frequency information of signal, but not
It can determine the time that these frequency signals occur, the localised information of time caused to be lost, high frequency and low can not be embodied simultaneously
The characteristic of frequency, there are limitations；(2) feature extraction being carried out to non-intrusion type load monitoring with S-transformation method, detection accuracy is higher,
Classify relatively accurate, but the operand of S-transformation is larger, real-time is difficult to ensure；(3) wavelet transformation can preferably be positioned at identification
The wave character of signal, has the characteristics that multiresolution, all has the ability of characterization signal local feature in time domain and frequency domain, but
The real small echo arithmetic speed of the tradition of use is slow, and time-consuming.
When the feature vector that numerous pairs load is identified with different algorithms, generally require to spend a large amount of calculating
Resource is suitble to the input data model of the load to construct.However in the case where data do not have class label, there is the machine of supervision
Device learning algorithm generally can not predict their label, this may cause the mistake in NILM application to one or more load
Classification, dependence manual intervention is strong, and practicability is not high.
Summary of the invention
Present invention is primarily aimed at provided a kind of non-intrusion type load identification method, dug based on Design of Wavelet and data
Digging coorinated training to improve identification accuracy reduces the computational complexity of remained capacity, while providing a kind of test macro to this
Recognition methods is tested.
The present invention is achieved through the following technical solutions:
A kind of non-intrusion type load recognition methods based on coorinated training, is used for electricity system, and the electricity system includes
Several loads, comprising:
It captures the electricity system and loads the transient current signal after switching every time, obtain transient current signal data collection；
Several Wavelet Clusters are obtained by Design of Wavelet and cluster；
To each transient current signal, the covariance of each small echo Yu the transient current signal is calculated, it will be with the wink
State current signal has small echo of the small echo of maximum covariance as the load；
The wavelet coefficient energy that the load is calculated by wavelet transform, the feature vector as the load；
According to the feature vector of each load, by algorithm more than a pair, in conjunction with decision-tree model and k nearest neighbor classifier to institute
It states data set and carries out coorinated training, to classify to each transient current signal in the data set, and predict accordingly each negative
The classification of load.
Further, the progress coorinated training includes:
Based on transient current signal one k nearest neighbor classifier of training；
Original load class is predicted based on the k nearest neighbor classifier；
The decision-tree model for predicting final load class is determined according to original loadtype, and passes through the decision tree
The final load class of model prediction.
Further, when predicting final load class by the decision-tree model, best point is determined using Gini coefficient
Cutpoint, and final load class is predicted accordingly.
Further, when obtaining several Wavelet Clusters by Design of Wavelet and cluster, based on standardization, two-track position orthogonality and
Low-pass filtering criterion carries out cluster operation by k-means clustering algorithm to design small echo.
Further, the length of the small echo is 6, and number is 91.
A kind of test macro, including electricity system, exchange power calibration, current transformer and load identifier, the use
Electric system includes several different types of loads, and the exchange power calibration is used to be each load supplying, the load identifier
Including data acquisition module and load identification module, the data acquisition module connects the current transformer, passes through the electricity
Current transformer captures the electricity system and loads the transient current signal after switching every time, and the load identification module is for executing
Step in load recognition methods as described above.
Further, the electricity system includes the load of incandescent lamp, four seed type of energy-saving lamp, computer and charger, respectively
Load is connect by electronic switch with the exchange power calibration.
Compared with prior art, non-intrusion type load recognition methods provided by the invention is dug based on Design of Wavelet and data
Dig coorinated training, by algorithm more than a pair, in conjunction with decision-tree model and k nearest neighbor classifier to transient current signal data collection into
Row coorinated training to classify to each transient current signal that data are concentrated, and predicts the classification of each load accordingly, thus real
Showed it is following the utility model has the advantages that
1. wavelet transformation can preferably position the wave character of signal to be identified, has the characteristics that multiresolution, in time domain
All there is the ability of characterization signal local feature with frequency domain, be conducive to extract the effective characteristic information of load.
2. the algorithm based on k nearest neighbor classifier and decision-tree model coorinated training not only reduces the complexity of calculating,
And improve precision of prediction.
3. the recognition methods has the calculating time short, the advantage of high reliablity with the increase of load number.
Detailed description of the invention
Fig. 1 a is small echo cluster assessment figure (square error)；
Fig. 1 b is small echo cluster assessment figure (silhouette coefficient)；
Fig. 2 is test principle schematic diagram；
Fig. 3 a is energy-saving lamp current waveform figure；
Fig. 3 b is incandescent lamp current waveform figure；
Fig. 3 c is charger current waveform diagram；
Fig. 3 d is computer current waveform figure；
Fig. 4 is the coorinated training flow diagram based on k nearest neighbor classifier and decision tree；
Fig. 5 is the whole nicety of grading contrast schematic diagram of OAR method and coorinated training.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below with reference to embodiment and attached drawing, to this
Invention is described in further detail.
Non-intrusion type load recognition methods provided by the invention based on coorinated training is used for electricity system, electricity system
Including several loads.As shown in figure 4, recognition methods includes step S1, S2, S3, S4, S5.
Step S1: capture electricity system loads the transient current signal after switching every time, obtains transient current signal data
Collection.
A test macro can be designed to verify the recognition methods, as shown in Fig. 2, the test macro includes electricity consumption system
System, exchange power calibration 5, current transformer 6 and load identifier, electricity system include several different types of loads, exchange
Power calibration 5 is used to be each load supplying, and load identifier includes data acquisition module 7 and load identification module 8, data acquisition
Module 7 connects current transformer 6, captures the transient current signal after electricity system loads switching every time by current transformer 6,
Load identification module 8 is used to execute remaining step in above-mentioned load recognition methods.In the test macro, electricity system includes white
The load of vehement lamp 1,4 four seed type of energy-saving lamp 2, computer 3 and charger, each load by electronic switch with exchange power calibration 5
Connection.
It is tested by building PSCAD/EMTDC emulation platform.Simulation time is set as 10 seconds, is cut after being supported on 3 seconds
It changes.Current signal is sampled with 256 sampled points of each period, and the time in each period is 0.0167 second.It is cut using load
The difference in two periods after alternatively is classified.Data collection is carried out by simulation switching different loads, every kind of load can
To be opened in different combinations.8 kinds of switch states when the load switching of energy-saving lamp 2 are listed shown in following table (table 1), by cutting
Changing 4 loads can produce 32 kinds of different situations, and combination will be repeated 5 times every time.
Logical zero indicates that switch is in an off state in upper table, and logical one indicates closed state.Jumping symbol indicates negative
Load switches over.Setting voltage changes from 210V to 230V, the robustness using 2V as step-length testing algorithm under different voltages.Reason
In the case of thinking, experiment can generate 1760 different data records (× 11 voltage steps of 32 No. × 5 switches of combination).Record
Each transition of the handover event within two periods, is the pre-switch of preceding half period and a half period after switching respectively,
To ensure to capture the entire transient state period, the current temporary state signal of 4 loads is as shown in Fig. 3 a, Fig. 3 b, Fig. 3 c and Fig. 3 d.Use k-
Data set is divided into training set and test set by fo l d cross validation, wherein 90% is used to train, 10% for testing.
Step S2: several Wavelet Clusters are obtained by Design of Wavelet and cluster.To Design of Wavelet in step S2 and cluster difference
It is described as follows.
Design of Wavelet:
Wavelet transformation can preferably retain and extract the time-frequency details characteristic of signal as a kind of time-frequency analysis technology, wide
General to apply in power system failure diagnostic, on-line monitoring field, wavelet transformation can pass through a scaling function V and a small echo letter
Number W is analyzed, such as formula (1) and (2)
In order to select a specific small echo to each load, need to create the small echo list that can be used.Pass through
When Design of Wavelet and cluster obtain several Wavelet Clusters, according to the three of small echo features (standardization, two-track position orthogonality and low pass filtered
Wave), small echo is designed based on standardization, two-track position orthogonality and low-pass filtering criterion.
1) it standardizes:
2) two-track position orthogonality:
3) low-pass characteristic:
It is finite response according to three equation low-pass filters above, the length of filter can be indicated with n, and
And the quantity of freedom degree can be indicated with n/2-1.Since the filter that length is 6 can produce a large amount of small echo and echo signal
It is matched, so directly taking n=6, then can have six different equations, two freedom degrees.By the way that k=1 to 6 is substituted into
(3)-(5) formula solves six low-pass filter coefficients, obtains following (6) equation
For indicating two freedom degrees, c is indicated a and b with a and b in formula.In order to calculate wavelet coefficient, a and b can be from-π
It is changed to π according to step-length for 0.1 π.This is 21 different parameters of a and b assignment, can produce 441 small echos therewith
Echo signal matching afterwards.
Wave Cluster:
In order to reduce the computation burden during late design, cluster operation is carried out by k-means clustering algorithm, to production
441 raw small echos are grouped the small echo with same shape using clustering algorithm, the specific steps are as follows:
1) cluster of k-means: using k means clustering method, by calculating the Euclidean distance between each small echo pair
Δλ,vDetermined.Two wavelet function ψλ(l) and ψv(l) their Euclidean distance Δλ,vIt is calculated by formula (7):
Remember and share k cluster, cluster center is respectively u1, and u2, u3 ... ..uk, the number of samples of each cluster is N1, N2, N3 ...
..NK.Use square error formula (8) as objective function, reduces the square error.
In formulaIt is the mass center of j-th of cluster, ψiIndicate i-th of data object in j-th of cluster.By increasing number of clusters
Square error can be reduced, but number of clusters is more, the burden of subsequent calculating is heavier.In order to solve this contradiction, wheel is introduced
Wide coefficient (Silhouette Coefficient) assesses Clustering Effect, selects representative small echo, judges that k mean value is poly-
The cluster number of class, calculation formula such as formula (9).
Si=(bi-ai)/max(ai-bi) (9)
A in formulaiIt is the average distance of vector i every other object into cluster, and biIt is vector i to all non-where itself
The average distance of remaining point of cluster, it is seen that the value of silhouette coefficient is more to level off between [- 1,1] and 1 represent cohesion degree and separating degree
It is all relatively excellent.
By Design of Wavelet and cluster, the length of the small echo finally obtained is 6, and number is 91.Fig. 1 a and Fig. 1 b difference
Illustrate the square error summation and silhouette coefficient value of the small echo that the filter for the use of length being 6 generates.Show in figure when small
When the quantity of wave cluster is 91, maximum Silhouette coefficient is that 1, SSE is 4.825 × 10-29.
Step S3: to each transient current signal, calculating the covariance of each small echo Yu transient current signal, will be with transient state
Current signal has small echo of the small echo of maximum covariance as load.The analysis of covariance:
Herein, covariance will be used to judge, and will select the echo signal X generated with each loaddWith most
The small echo of big covariance, the transient signal that echo signal can be generated by switching load are collected.Then formula (10) are used
To calculate the covariance of each small echo and echo signal.There to be the small echo of maximum covariance as the small echo of the load.
L indicates the sampling instant (l=1,2,3 ... L) of signal in formula,WithIt is the mean value of echo signal Xd respectively
With representative small echo ψRMean value.
Each of the current transient waveform of capture and 91 small echos are compared, and calculate the association side of each small echo
Difference will have the small echo of maximum covariance as the small echo in the load matched, and it is small that 4 fictitious loads produce 4 representativenesses
Wave.Following table (table 2) shows the small echo serial number after each load matched.
Step S4: the feature vector by the wavelet coefficient energy of wavelet transform computational load, as load.The step
Feature identification is carried out using the energy of wavelet coefficient in rapid:
For NILM application in, need to extract load switching after echo signal X in include hiding feature, can by pair
Echo signal XdWavelet coefficient energy as extracting.Use echo signal XdVariation and scaling function V convolution fortune
It is cA that calculation, which obtains low frequency component coefficient,j(l), echo signal XdVariation and the convolution algorithm of wavelet function W obtain high fdrequency component system
Number cDj(l)。
cAj(l)=< Xd,Vj,l>,cDj(l)=< Xd,Wj,l> (11)
The energy of each wavelet coefficient can be calculated by formula (11), which can be used as between different loads
Characteristic quantity distinguishes.
Step S5: according to the feature vector of each load, by algorithm more than a pair, in conjunction with decision-tree model and k arest neighbors point
Class device carries out coorinated training to data set, to classify to each transient current signal that data are concentrated, and predicts accordingly each negative
The classification of load.The step is trained data set using the thought of decision-tree model and k nearest neighbor classifier coorinated training.It should
Process firstly generates two datasets: the data L of label and unlabelled data U；The data set of label is divided into two
A subset LAAnd LB, developing two disaggregated models by decision tree and NN training is respectively hAAnd hB；After data set generation,
Subset U' is obtained by randomly choosing u record from unlabelled data set U, is sent to hAAnd hBTwo disaggregated models are used
In Tag Estimation.If with tape label data train come model the result of no Tag Estimation is consistent, thus can
Trained model before is further trained with the data set of no label, until unlabelled data all in U are used,
Coorinated training is just completed at last.
The advantage of binary classifier is not only utilized in the classification method of coorinated training, while also increasing the multiple classifications of processing
The ability of label.It combines a k nearest neighbor classifier and decision-tree model to optimize by using OAR method, this will be obtained
The advantage of two kinds of sorting algorithms.
OAR method: by creating four individual decision-tree models, by algorithm more than a pair (One Against Rest,
OAR) classify.Method more than a pair is a series of two classification devices of construction, and a classification and remaining class are separated, and is created
Decision tree.For example a kind of label of energy-saving lamp 2 is positive, remaining load class, which is labeled, to be negative.Then using the method for DWT to all
Transient current signal carries out wavelet transformation, and the energy value of small echo is used as training set.Need to calculate the wavelet coefficient of four loads
Energy value, the energy of coefficient is sent in their own decision-tree model, then different loads classification is positive class and negative
Class.Ultimately form the vector comprising four classifiers.Classified using OAR method, using Monte Carlo method to classifier
It tests.It is recognised that being just enough to obtain consistent result by 1000 Monte Carlo iteration from experiment test.As a result
It see the table below (table 3):
Nicety of grading is calculated using formula (13):
TP indicates the positive class event correctly classified, and TN is the negative class event correctly classified, and P is total positive class event, and N is total
Negative class event.
The nicety of grading of energy-saving lamp 2 and charger 4 is significant lower in table 3 known to observation.In order to find out the decision for using OAR
Tree-model there are the shortcomings that, we introduce two indices and analyze, and are the success rate (TPR) to positive class event category respectively
With the success rate (TNR) to negative class event category, their own classification accuracy such as formula (14)-(15) are calculated:
Shown in recognition accuracy such as following table (table 4):
It can see from table 4, when obtaining precision using TPR, it can be seen that the precision of remained capacity is higher than OAR algorithm
The precision of itself illustrates that decision-tree model is higher to the correctness of positive class event category.But it can see pair from the accuracy of TNR
The accuracy of identification of energy-saving lamp 2 and charger 4 is all very low.This illustrates that decision-tree model is classified existing defects in negative class event.
In order to solve this problem, we introduce the method for coorinated training to mitigate decision-tree model to the pressure of negative class event handling
Power.
Unlike OAR method, coorinated training does not pass through the energy coefficient training decision-tree model of small echo, base
A k nearest neighbor classifier is first trained in transient current signal itself, k value is set as 5 neighbours by repetition test.It will train
K nearest neighbor classifier as original classification device, predict original load class.Then it is determined according to original loadtype for pre-
The decision-tree model of final load class is surveyed, and final load class is predicted by decision-tree model.It is pre- by decision-tree model
When surveying final load class, optimal partition point is determined using Gini coefficient, and predict final load class accordingly.
Classify by this way to new echo signal, k nearest neighbor classifier, primary discrete small is only utilized
Wave conversion (DWT) and a decision tree classifier carry out coorinated training and classify.With use four DWT and four decision tree classifications
The OAR method that device is classified compares.Coorinated training algorithm not only reduces the complexity of calculating and improves precision of prediction.
Accuracy computation formula in formula (13) is equally used for calculating the classification accuracy of the decision-tree model of collaboration test.
Due to classification method be changed into the method for coorinated training according to OAR method, so accuracy measurement mode keep identical, and
And the method for obtaining precision also remains unchanged.It compared classifying using coorinated training method and tri- kinds of TPR, TNR in following table (table 5)
Situation precise manner obtained:
From the point of view of the nicety of grading of table 5, compared with the OAR method in table 3, coorinated training method generates preferably classification knot
Fruit.TPR precision shown in table 5 shows that the precision of decision-tree model itself is not improved, this is because building decision-tree model
Method there is no changing, therefore the nicety of grading of positive class event is not improved.But it can be seen that from the TNR precision in table 5
The error rate of TN model has almost completely disappeared.The precision difference very little of two TPR results, shows that k is nearest in table 5 and table 4
The original tag classification of adjacent classifier directly affects the accuracy of cooperation detection classifier.However, since decision-tree model has
There is the advantage of precision 99%, the significant decline of mistake classification results causes the accuracy of whole classifier to improve.Fig. 5 is shown
Comparison in two methods entirety niceties of grading.
Compare the computational complexity of every kind of method.In OAR method, a unique small echo is selected for each load, it is small
The quantity of wave conversion is equal to the quantity loaded in system.Then classify to each load.It is surveyed in this simple four load
In examination, about 10-12 milliseconds is run in the single thread of PC.And in coorinated training method, each load selection one is unique
Small echo, it is only necessary to execute a wavelet transformation, the quantity no matter loaded all only needs to carry out double classification, therefore cooperates with
The trained method testing time has only used 6ms.It may seem very little in the example that this difference may be loaded at four, but such as
The quantity of fruit load increases more, and this method, which will have, calculates temporal advantage.It is calculated this paper presents a kind of using decision tree
Method and the non-intrusion type load of nearest neighbor algorithm coorinated training know method for distinguishing.Producing one group of length by Design of Wavelet first is
6 small echo, by cluster after can be by 91 different small wave components.Then circuit evolving 1760 are built using PSCAD emulation
The data set of transient signal composition selects the small of tight fit using the analysis of covariance to the transient signal of each load
Wave.The type of feature extraction load is carried out by calculating the energy of wavelet coefficient using the small echo matched.Finally, herein to mesh
(OAR) sorting algorithm and k nearest neighbor classifier and decision Tree algorithms coorinated training compare and obtain knot more than preceding common a pair
By.The result shows that by using coorinated training method, the classification error that occurs when can obviously eliminate using OAR classification method.
The accuracy integrally classified is dramatically increased, the defect to true negative class event category accuracy is overcome, OAR arithmetic accuracy is
74%, the accuracy of identification of coorinated training is 99.5%.This explanation, the method for coorinated training have bright in the precision integrally classified
Aobvious advantage.
Above-described embodiment is only preferred embodiment, the protection scope being not intended to limit the invention, in spirit of the invention
With any modifications, equivalent replacements, and improvements made within principle etc., should all be included in the protection scope of the present invention.
Claims (7)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810989105.2A CN109165604A (en) | 2018-08-28 | 2018-08-28 | The recognition methods of non-intrusion type load and its test macro based on coorinated training |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810989105.2A CN109165604A (en) | 2018-08-28 | 2018-08-28 | The recognition methods of non-intrusion type load and its test macro based on coorinated training |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109165604A true CN109165604A (en) | 2019-01-08 |
Family
ID=64893076
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810989105.2A Pending CN109165604A (en) | 2018-08-28 | 2018-08-28 | The recognition methods of non-intrusion type load and its test macro based on coorinated training |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109165604A (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109387712A (en) * | 2018-10-09 | 2019-02-26 | 厦门理工学院 | Non-intrusion type cutting load testing and decomposition method based on state matrix decision tree |
| CN110188437A (en) * | 2019-05-22 | 2019-08-30 | 上海人赢网络科技有限公司 | A kind of ready-made clothes size generation method and its device based on machine learning |
| CN110244149A (en) * | 2019-07-05 | 2019-09-17 | 四川长虹电器股份有限公司 | Non-intrusion type electrical appliance recognition based on current amplitude standard deviation |
| CN110244150A (en) * | 2019-07-05 | 2019-09-17 | 四川长虹电器股份有限公司 | A kind of non-intrusion type electrical appliance recognition based on root mean square and standard deviation |
| CN110443302A (en) * | 2019-08-02 | 2019-11-12 | 天津相和电气科技有限公司 | Load discrimination method and its application based on Fusion Features and deep learning |
| CN111161097A (en) * | 2019-12-31 | 2020-05-15 | 四川大学 | Method and device for detecting switch event based on event detection algorithm of hypothesis test |
| CN111259869A (en) * | 2020-03-11 | 2020-06-09 | 北京慧飒科技有限责任公司 | Non-invasive electrical cluster load fault identification method |
| CN111275069A (en) * | 2019-12-10 | 2020-06-12 | 深圳供电局有限公司 | Non-invasive load monitoring method |
| CN111898694A (en) * | 2020-08-07 | 2020-11-06 | 广东电网有限责任公司计量中心 | Non-invasive load identification method and device based on random tree classification |
| CN113673480A (en) * | 2021-09-03 | 2021-11-19 | 武汉盛帆电子股份有限公司 | Non-intrusive load identification method and device |
| CN113899944A (en) * | 2021-09-30 | 2022-01-07 | 广东电网有限责任公司 | Detection method and device for power load switching point |
| CN114970648A (en) * | 2021-02-20 | 2022-08-30 | 青岛特来电新能源科技有限公司 | New energy equipment identification method, device and medium |
| CN115407157A (en) * | 2022-10-31 | 2022-11-29 | 南昌工程学院 | Complex multi-state electrical appliance load event distinguishing method and device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160035052A1 (en) * | 2010-08-30 | 2016-02-04 | Bao Tran | Systems and methods for reducing energy usage |
| CN106532694A (en) * | 2016-12-20 | 2017-03-22 | 武汉大学 | Distributed non-invasive load measurement system |
| CN106680621A (en) * | 2016-12-02 | 2017-05-17 | 华北电力大学 | A Residential Load Electricity Identification Method Based on Current Signal Separation |
| CN107330517A (en) * | 2017-06-14 | 2017-11-07 | 华北电力大学 | One kind is based on S_Kohonen non-intrusion type resident load recognition methods |
| CN108054749A (en) * | 2017-11-27 | 2018-05-18 | 华北电力大学 | A kind of non-intrusive electrical load decomposition method and device |
-
2018
- 2018-08-28 CN CN201810989105.2A patent/CN109165604A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160035052A1 (en) * | 2010-08-30 | 2016-02-04 | Bao Tran | Systems and methods for reducing energy usage |
| CN106680621A (en) * | 2016-12-02 | 2017-05-17 | 华北电力大学 | A Residential Load Electricity Identification Method Based on Current Signal Separation |
| CN106532694A (en) * | 2016-12-20 | 2017-03-22 | 武汉大学 | Distributed non-invasive load measurement system |
| CN107330517A (en) * | 2017-06-14 | 2017-11-07 | 华北电力大学 | One kind is based on S_Kohonen non-intrusion type resident load recognition methods |
| CN108054749A (en) * | 2017-11-27 | 2018-05-18 | 华北电力大学 | A kind of non-intrusive electrical load decomposition method and device |
Non-Patent Citations (3)
| Title |
|---|
| J.CHUNG 等: "Non-Intrusive Load Monitoring Using Wavelet Design and Co-Testing of Machine Learning Classifiers", 《2016 IEEE ELECTRICAL POWER AND ENERGY CONFERENCE》 * |
| JESSIE M. GILLIS 等: "Non-Intrusive Load Monitoring Using Semi-Supervised Machine Learning and Wavelet Design", 《IEEE TRANSACTIONS ON SMART GRID》 * |
| 王丹 等: "基于改进聚类和DTW的非侵入式负荷识别", 《信息通信》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109387712A (en) * | 2018-10-09 | 2019-02-26 | 厦门理工学院 | Non-intrusion type cutting load testing and decomposition method based on state matrix decision tree |
| CN110188437A (en) * | 2019-05-22 | 2019-08-30 | 上海人赢网络科技有限公司 | A kind of ready-made clothes size generation method and its device based on machine learning |
| CN110244149B (en) * | 2019-07-05 | 2021-03-16 | 四川长虹电器股份有限公司 | Non-invasive electrical appliance identification method based on current amplitude standard deviation |
| CN110244149A (en) * | 2019-07-05 | 2019-09-17 | 四川长虹电器股份有限公司 | Non-intrusion type electrical appliance recognition based on current amplitude standard deviation |
| CN110244150A (en) * | 2019-07-05 | 2019-09-17 | 四川长虹电器股份有限公司 | A kind of non-intrusion type electrical appliance recognition based on root mean square and standard deviation |
| CN110443302A (en) * | 2019-08-02 | 2019-11-12 | 天津相和电气科技有限公司 | Load discrimination method and its application based on Fusion Features and deep learning |
| CN110443302B (en) * | 2019-08-02 | 2023-06-09 | 天津相和电气科技有限公司 | Load identification method based on feature fusion and deep learning and application thereof |
| CN111275069B (en) * | 2019-12-10 | 2023-11-28 | 深圳供电局有限公司 | A non-invasive load monitoring method |
| CN111275069A (en) * | 2019-12-10 | 2020-06-12 | 深圳供电局有限公司 | Non-invasive load monitoring method |
| CN111161097A (en) * | 2019-12-31 | 2020-05-15 | 四川大学 | Method and device for detecting switch event based on event detection algorithm of hypothesis test |
| CN111161097B (en) * | 2019-12-31 | 2023-07-25 | 四川大学 | Method and device for detecting switching events based on hypothesis testing event detection algorithm |
| CN111259869A (en) * | 2020-03-11 | 2020-06-09 | 北京慧飒科技有限责任公司 | Non-invasive electrical cluster load fault identification method |
| CN111898694B (en) * | 2020-08-07 | 2021-09-17 | 广东电网有限责任公司计量中心 | Non-invasive load identification method and device based on random tree classification |
| CN111898694A (en) * | 2020-08-07 | 2020-11-06 | 广东电网有限责任公司计量中心 | Non-invasive load identification method and device based on random tree classification |
| CN114970648A (en) * | 2021-02-20 | 2022-08-30 | 青岛特来电新能源科技有限公司 | New energy equipment identification method, device and medium |
| CN113673480A (en) * | 2021-09-03 | 2021-11-19 | 武汉盛帆电子股份有限公司 | Non-intrusive load identification method and device |
| CN113673480B (en) * | 2021-09-03 | 2024-08-02 | 武汉盛帆电子股份有限公司 | Non-intervention type load identification method and device |
| CN113899944A (en) * | 2021-09-30 | 2022-01-07 | 广东电网有限责任公司 | Detection method and device for power load switching point |
| CN113899944B (en) * | 2021-09-30 | 2023-11-10 | 广东电网有限责任公司 | Detection method and device for power load switching point |
| CN115407157A (en) * | 2022-10-31 | 2022-11-29 | 南昌工程学院 | Complex multi-state electrical appliance load event distinguishing method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109165604A (en) | The recognition methods of non-intrusion type load and its test macro based on coorinated training | |
| CN111830347B (en) | Two-stage non-invasive load monitoring method based on event | |
| CN109146705B (en) | Method for detecting electricity stealing by using electricity characteristic index dimension reduction and extreme learning machine algorithm | |
| CN109856299A (en) | A kind of transformer online monitoring differentiation threshold value dynamic setting method, system | |
| Saini et al. | Classification of power quality events–a review | |
| Lin et al. | Detection and classification of multiple power-quality disturbances with wavelet multiclass SVM | |
| CN108616120B (en) | Non-invasive power load decomposition method based on RBF neural network | |
| CN105974265B (en) | A kind of electric network fault cause diagnosis method based on svm classifier technology | |
| CN102520341B (en) | Analog circuit fault diagnosis method based on Bayes-KFCM (Kernelized Fuzzy C-Means) algorithm | |
| Jiang et al. | Spatial-temporal synchrophasor data characterization and analytics in smart grid fault detection, identification, and impact causal analysis | |
| CN111444940A (en) | Fault diagnosis method for critical parts of fan | |
| CN111368904B (en) | Electrical equipment identification method based on electric power fingerprint | |
| CN109635928A (en) | A kind of voltage sag reason recognition methods based on deep learning Model Fusion | |
| CN104809473B (en) | A kind of classification method towards electric power online acquisition data | |
| CN108562821B (en) | Method and system for determining single-phase earth fault line selection of power distribution network based on Softmax | |
| CN109389325A (en) | Transformer substation electronic transducer state evaluating method based on wavelet neural network | |
| CN114091549A (en) | Equipment fault diagnosis method based on deep residual error network | |
| CN118332424A (en) | Small sample cross-domain classification prediction method for electromagnetic compatibility defects of battery simulator | |
| CN114236234A (en) | A method for electrical feature identification based on fundamental and harmonic hybrid criteria | |
| CN115712871A (en) | Power electronic system fault diagnosis method combining resampling and integrated learning | |
| CN111275204A (en) | A Transformer State Recognition Method Based on Hybrid Sampling and Ensemble Learning | |
| CN115169405A (en) | Hotel guest room equipment fault diagnosis method and system based on support vector machine | |
| CN118133190B (en) | Load identification model construction method and load identification method based on BN relation network | |
| CN119203718A (en) | Adaptive multi-scale transient voltage stability assessment method based on improved SK-BiLSTM | |
| He et al. | Application and comparative analysis of traditional machine learning and deep learning in transmission line fault classification |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190108 |