CN113947159A - Real-time online monitoring and identification method for electrical load - Google Patents

Real-time online monitoring and identification method for electrical load Download PDF

Info

Publication number
CN113947159A
CN113947159A CN202111247143.9A CN202111247143A CN113947159A CN 113947159 A CN113947159 A CN 113947159A CN 202111247143 A CN202111247143 A CN 202111247143A CN 113947159 A CN113947159 A CN 113947159A
Authority
CN
China
Prior art keywords
real
data
algorithm
ent
combination
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202111247143.9A
Other languages
Chinese (zh)
Other versions
CN113947159B (en
Inventor
魏广芬
赵航
胡春华
刘骞
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.)
Shandong Technology and Business University
Yantai Dongfang Wisdom Electric Co Ltd
Original Assignee
Shandong Technology and Business University
Yantai Dongfang Wisdom Electric Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shandong Technology and Business University, Yantai Dongfang Wisdom Electric Co Ltd filed Critical Shandong Technology and Business University
Priority to CN202111247143.9A priority Critical patent/CN113947159B/en
Publication of CN113947159A publication Critical patent/CN113947159A/en
Application granted granted Critical
Publication of CN113947159B publication Critical patent/CN113947159B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Mathematical Analysis (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Optimization (AREA)
  • Computational Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Databases & Information Systems (AREA)
  • Algebra (AREA)
  • Probability & Statistics with Applications (AREA)
  • Operations Research (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a real-time online monitoring and identifying system for an electricity load, which identifies the type and the working state of the electricity load by utilizing a Shannon entropy weighted voting algorithm model, wherein the determination process of a base classifier combination in the Shannon entropy weighted voting algorithm model is as follows: selecting E base classifiers to carry out
Figure DDA0003321448620000011
Combining to obtain a plurality of different algorithm model combinations, calculating the diversity Ent of the different algorithm model combinations, and determining the optimal value range Ent of Ent*(ii) a For the object belonging to Ent*The combination of algorithm models in the range, and the Shannon entropy of the combination of algorithm models and the addition of the entropyThe accuracy of the weight vote, denoted ACCse，Ent*Inner distance Ent*ACC with closest intermediate point and largest numerical valueseAnd the corresponding algorithm model combination is the optimal learner combination which is marked as Com, and the trained Shannon entropy weighted voting algorithm model is obtained. Therefore, the real-time online monitoring and identifying system for the power load has less hardware investment and high identifying efficiency and accuracy.

Description

Real-time online monitoring and identification method for electrical load
Technical Field
The invention relates to the field of electric power information acquisition, in particular to a real-time online monitoring and identifying method for an electric load.
Background
With the proposition of the strategy of 'carbon peak reaching and carbon neutralization', the environment-friendly, low-carbon, energy-saving and emission-reduction become normal and mainstream, wherein the optimized management of the electrical load of residents has larger energy-saving and emission-reduction potential.
The relevant detailed information of the household appliance load is obtained through the load monitoring technology, so that the method is not only beneficial to optimizing the operation and management of a power grid of a power company, but also beneficial to efficiently managing power utilization equipment by resident users, and therefore powerful support is provided for energy conservation and emission reduction. According to the Monitoring method, the current Load Monitoring technology mainly includes two types, Non-Intrusive Load Monitoring (NILM) and Intrusive Load Monitoring (ILM). The traditional load monitoring method generally records the use condition by installing a sensor on the electrical equipment of a user, belongs to an intrusive mode, needs to add a module on the electrical equipment, is high in economic cost and is not beneficial to popularization and application. In comparison, the non-intrusive load monitoring does not increase marginal cost, and generally, a monitoring device is installed at an intelligent electric meter, so that energy consumption data of the whole family is obtained, and the operation state of each household electric device is accurately identified by using methods such as event detection, mode identification and the like. The method has the advantages of good operability, low implementation cost, high user acceptance, easy popularization and wide prospect.
At present, in order to meet the requirement of real-time load online identification, an online identification system capable of processing data in real time is needed, the system needs to receive and store the data in real time, and meanwhile, the waveform, the real-time calculation characteristics and the loading model are displayed in a display area in real time to judge the category of an electric appliance in real time. Therefore, multiple threads are needed to realize the functions, so that the system is not jammed, the real-time requirement is met, and the real-time online monitoring and identification tasks are completed.
Compared with the traditional single learning algorithm, the integrated learning method has stronger advantages, and Boosting, Bagging, Stacking, Voting and other methods are formed at present. The Voting method is an integration method commonly used in classification problems, and comprises an absolute majority Voting method, a relative majority Voting method, a weighted Voting method and the like. The method effectively synthesizes the prediction results of various base classifiers, and the classification and identification effects are outstanding. However, in the experimental process of constructing a strong learner with high accuracy and strong generalization ability by using a plurality of weak heterogeneous base classifiers, the following problems are found to occur in the application of the Voting algorithm:
(1) the relative majority voting method only needs to acquire the option with the largest number of votes as the final result, but when the majority classifier predicts an error, the integrated model prediction result is also wrong, so that the recognition accuracy after integration is low.
(2) According to different scenes, the prediction accuracy of various base classifiers is different, namely the prediction results of the base classifiers have differences. Therefore, different base classifiers are selected to vote, the obtained voting results are inconsistent, and sometimes the combined voting results of the selected base classifiers are not as good as those of a single learner. How to select the base classifier in the integration method is an urgent problem to be solved.
(3) The weighted voting method is to configure different weights for each base classifier, and when an appropriate weight is given, the result obtained by the weighted voting method can be better than that obtained by the base classifier, and the result obtained by the weighted voting method is also better than that obtained by the relative majority voting method.
There is therefore a need for improvements in the prior art.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: an electric load real-time online monitoring and identifying method is provided.
In order to solve the technical problems, the technical scheme adopted by the invention is as follows:
a real-time online monitoring and identification method for electrical loads comprises the following steps:
s1, receiving an original signal sent by the intelligent electric meter by using the main thread, and analyzing according to a protocol to obtain original data;
s2, storing the original data by the sub thread 1;
s3, the sub-thread 2 displays the original data in real time through a display device;
s4, analyzing the voltage and current data of each period in the original data by the main thread, and extracting characteristic data;
s5, storing the characteristic data by the sub-thread 3, and storing the characteristic data to the local in real time;
s6, calling a display function by a sub-thread 4 to initialize a waveform diagram of the electric appliance to be displayed, wherein the waveform diagram displays the electricity utilization characteristic data of the electric appliance;
s7, the thread 5 reads the real-time feature data and transmits the feature data to the display function in the thread 4, and the feature data in the display is updated;
s8, detecting the event by the thread 4 by using an event detection algorithm, and detecting whether the event occurs according to the real-time characteristic data in the thread 5; the event means that the change value of the characteristic data is higher than a set change threshold value in a set detection period;
s9, when an event occurs, loading the trained Shannon entropy weighting voting algorithm model by using the sub-thread 6;
and S10, reading the real-time characteristic data by the sub-thread 6, and inputting the real-time characteristic data into the Shannon entropy weighting voting algorithm model to realize the online identification of the type and the working state of the electric appliance.
Compared with the prior art, the invention has the following technical effects:
and the system is ensured not to be jammed by running multiple threads, the real-time requirement is met, and the real-time online monitoring and identification tasks are completed. And realizing the online identification of the category and the working state of the electric appliance by utilizing a Shannon entropy weighting voting algorithm model.
On the basis of the technical scheme, the invention can be further improved as follows.
Preferably, the display device in step S2 presents the original data in a text format.
Preferably, the waveforms in step S6 include voltage waveform, current waveform, active power waveform, reactive power waveform, V-I trace and harmonic amplitude;
preferably, the characteristic data in step S4 includes active power (P), reactive power (Q), apparent power (S), power factor (λ), (2g-1) subharmonic, where g is a positive integer, such as 1 subharmonic (har1), 3 subharmonic (har3), …, 63 subharmonic (har 63);
preferably, in step S5, the shannon entropy weighted voting algorithm model is trained as follows,
A1、
the Feature data is stored locally to form a Feature database, the main thread pair includes a Linear Discriminant Analysis (LDA), a Naive Bayes classifier (NB), a K-Nearest Neighbor algorithm (KNN), a Decision Tree classifier (DT), a Support Vector Machine (SVM), a Logistic Regression algorithm (LR), a feedback Neural Network (BPNN), and a Recursive Feature Elimination (Recursive Feature aggregation, RFE) to select effective Feature data of each algorithm, and the effective Feature data is used for training the seven algorithms to obtain a post-basis algorithm model of each algorithm, subsequently train a basis classifier, and train numbers of 7 basis classifiers from 0 to 6.
A2、
To 7 kinds of base classifiers
Figure BDA0003321448600000031
Combining to obtain 120 different algorithm model combinations, calculating the diversity Ent of various combinations by using a formula (1),
Figure BDA0003321448600000032
ent is diversity in formula (1), and m is a certain algorithm modeThe number of base classifiers in the type combination corresponds to one Ent, x in each algorithm model combinationkIs the kth sample in the feature data set X, K being the total number of samples in the feature data set X, z (X)k) Refers to a combination of m base classifiers for a sample xkThe number of correctly classified base classifiers can be obtained.
A3、
Calculating the optimal range Ent of the diversity according to the formula (3) by using the maximum value and the minimum value of the diversity*
Figure BDA0003321448600000033
In the formula (3), σ is Ent*The half interval range of (1) is an empirical set value, and is generally between 0 and 1;
for the case of selecting E base classifiers, the predicted classification probabilities are weighted and fused according to formula (4) for the case of n classification classes
Figure BDA0003321448600000041
P in formula (4)En(xk) Represents the sample x for the feature datakThe E-th classifier predicts the posterior probability of the nth class. w is aiIs the ith base classifier to sample xkThe fusion weight of (a);
Figure BDA0003321448600000042
in the formula (5), the reaction mixture is,
Hi(xk) Representing feature-based data samples xkThe Shannon entropy of the ith classifier on all classification classes is calculated;
Figure BDA0003321448600000043
j represents the classification category and i is the number of the base classifier.
A4、
Adding all the results calculated by the formula (6) to realize the posterior probability weighted fusion of the categories to obtain P' (x)k) As shown in formula (7)
Figure BDA0003321448600000044
It is based on data x in an algorithm model combinationkProbability of being judged as each category.
A5、
P' (x) as shown in equation (8)k) The column index of the element with the largest numerical value is for sample xkPrediction Class (x) ofk)，
Class(xk)＝argmax(P″(xk)) (8)
arg represents the sequence corresponding to the element with the largest value in the solution data list. Thereby obtaining the prediction results of all training samples.
A6:
Calculating the prediction accuracy ACC of each algorithm model combination based on the prediction results and the real results of all training samplesse
Figure BDA0003321448600000051
TPse、TNse、FNse、FPseThe number of the predicted result and the real result of the se combination on the training sample is shown as follows: TPseReally, the number of samples which show real results as positive classes is predicted as the positive classes; TN (twisted nematic)seTrue negative, which represents the number of samples with true results as negative class predicted as negative class; FPseIf the sample is false positive, the number of samples with true results of negative classes is predicted to be positive classes; FN (FN)seIf the true result is false negative, the number of samples with true results of positive class is predicted as negative class; c1 is a pairSelected E base classifiers
Figure BDA0003321448600000052
The number of algorithm model combinations obtained by combination.
A7: for the object belonging to Ent*Combining the algorithm models in the range, finding the value which is the largest and is at Ent*ACC closest to the middle position insideseAnd (4) corresponding algorithm model combination, namely the optimal learner combination which is marked as Com. The best learner combination Com is the trained Shannon entropy weighted voting algorithm model.
Further, the recursive feature elimination method uses a base model to perform multiple rounds of training, eliminates the features of a plurality of weight coefficients after each round of training, performs the next round of training based on a new feature set, and simultaneously uses 10-fold cross validation on the training set in each round of training to obtain the validation accuracy of the current feature combination. And after the training is finished, finding out the characteristic data combination which has the largest influence on the final prediction result. The most important characteristic data for the basic classification algorithm is screened out through the process, and the accuracy of the prediction result of the basic classification algorithm is improved.
Further, the method of calculating the active power (P), the reactive power (Q), the apparent power (S), the power factor (λ), the 1 st harmonic (har1), the 3 rd harmonic (har3), …, and the 63 rd harmonic (har63) is as follows:
firstly, the effective value of the voltage is calculated in the way shown in the formula (9).
Figure BDA0003321448600000053
Where N is the number of data points in the calculation cycle, UdVoltage values in a single data point;
the effective value of the current is calculated in the formula (10).
Figure BDA0003321448600000054
In the formula IdAs current values in a single data point
The active power P is equal to the average power of the N test points in the data segment, see equation (11).
Figure BDA0003321448600000061
Apparent power S equal to voltage effective value UrmsAnd the effective value of the current IrmsThe product of (c) is shown in equation (13).
S＝Urms×Irms (13)
The reactive power Q is calculated according to equation (12).
Figure BDA0003321448600000062
The power factor is usually expressed by λ, and is the ratio of the active power P of the ac circuit to the apparent power S, see equation (14).
λ＝P/S (14)
The collected current data are subjected to conversion calculation by combining an equation (15) and a fast Fourier transform algorithm to obtain 2n +1 harmonics, wherein n is an integer, 32 odd harmonics such as fundamental waves, 3 harmonics, 5 harmonics, … harmonics, 63 harmonics and the like are selected as characteristics and are respectively marked as Har1, Har3, Har5, … and Har63, and the corresponding harmonic frequencies are respectively 50Hz, 150Hz, 250Hz, … and 3.15 KHz.
Figure BDA0003321448600000063
In the formula (15), ω is 2 pi/T is the angular frequency of the periodic function, T is the period of the voltage, and corresponds to domestic alternating current in China for 0.02 second, wherein ω is 50Hz, and h is the harmonic frequency. a is0Is a direct current coefficient in the signal, ah，bhFor the AC coefficients in the signal, calculation is performed by Fourier transformAnd (6) obtaining.
Further, the optimal model combination obtained for the self-test data set comprises four base classifiers of LDA, NB, DT and LR.
Further, the best model combinations obtained for white (a global family and industry transient data set with a sampling frequency of 44.1KHz and a data set containing current and voltage data of more than 130 electrical) include four base classifiers KNN, DT, LR and BPNN.
The beneficial effect of adopting the further scheme is that the optimal model is combined, and then the optimal model in the combination is stored to the local so as to be called at any time in the using process of the system. Inputting real-time characteristic data into stored model, and generating prediction result in real time
Drawings
FIG. 1 is a flow diagram of the present invention.
FIG. 2 is a graph of the number of different algorithmic features when using measured data sets versus the accuracy of the validation set.
FIG. 3 is a graph of the number of different algorithmic features and the accuracy of a validation set when using a WHITED data set.
FIG. 4 is a flowchart of finding the optimal learner combination Com.
FIG. 5 is a graph showing the relationship between the average accuracy and the diversity when using the measured data set.
FIG. 6 is an enlarged partial view of FIG. 5;
figure 7 is the average accuracy and diversity relationship using the white dataset.
FIG. 8 is an enlarged partial view of FIG. 7;
figure 9 is a NILM frame diagram.
FIG. 10 is a diagram of characteristic data before and after an event occurs.
FIG. 11 is the accuracy of the LDA, NB, DT, LR and the algorithm herein testing on the measured data set.
FIG. 12 is the test accuracy of KNN, DT, LR, BPNN and the algorithm herein on a WHITED data set.
FIG. 13 is a test set confusion matrix for a self-test data set.
Fig. 14 is a real-time detection result of the hair dryer.
Detailed Description
The principles and features of this invention are described below in conjunction with the following drawings, which are set forth by way of illustration only and are not intended to limit the scope of the invention.
Example 1:
a multithreading-based non-intrusive load data real-time processing and identification method is used for carrying out real-time online monitoring and identification on the type and the running state of an electric appliance by acquiring real-time electric appliance electricity utilization data and comprises the following steps.
And step S1, receiving the original signal sent by the intelligent electric meter by using the main thread serial port, and analyzing according to a protocol to obtain original data.
In step S2, raw data is stored using child thread 1.
Step S3, using sub-thread 2 to display the raw data in real time to a display area of a display device, the display area displaying the raw data in a text format.
And step S4, the main thread analyzes the voltage and current data of each period in the original data, extracts features and obtains feature data.
In step S5, the sub-thread 3 is used to store the feature data, and the feature data is stored locally in real time.
Step S6, using sub thread 4 to call a real-time display function to initialize the oscillogram of the electric appliance to be displayed, wherein the wave form comprises a voltage wave form, a current wave form, an active power wave form, a reactive power wave form, a V-I track and a harmonic amplitude;
in step S7, the feature data is transmitted to the real-time display function in step S6 using the signal function of thread 5, so that the displayed feature data is updated and presented in the form of a waveform.
Step S8, the sub-thread 4 uses an event detection algorithm to detect events, and detects whether an event occurs according to the real-time characteristic data presented in step S7, wherein the event means that the change rate of the characteristic data is higher than a set change threshold value in a set detection period; the characteristic data used herein may be current data, and may also be active power, reactive power, etc.
And step S9, loading the trained Shannon entropy weighting voting algorithm model by using the sub-thread 6, reading the real-time characteristics, inputting the real-time characteristics into the Shannon entropy weighting voting algorithm model, and realizing the online identification of the type and the working state of the electrical appliance.
In the feature extraction described in step S4, the present invention extracts active power (P), reactive power (Q), apparent power (S), power factor (λ), 1-th harmonic (har1), 3-th harmonic (har3), …, 63-th harmonic (har63), and 36 features of 5 types in total.
The load characteristics are further divided into steady-state characteristics and transient characteristics, wherein the steady-state characteristics are electrical quantities of the load in a steady operation state, the transient characteristics are electrical quantities in a transient process, and the transient characteristics can be used for non-intrusive load identification alone or as supplementary characteristics of the steady-state characteristics.
Table 2 shows the typical characteristic data, and the respective advantages and disadvantages.
TABLE 2 common feature types
Figure BDA0003321448600000081
Active power (P), reactive power (Q), apparent power (S), power factor (λ), 1 th harmonic (har1), 3 rd harmonic (har3), …, 63 th harmonic (har63) are calculated.
Firstly, the effective value of the voltage is calculated in the way shown in the formula (9).
Figure BDA0003321448600000082
The effective value of the current is calculated in the formula (10).
Figure BDA0003321448600000083
The active power P is equal to the average power of the N test points in the data segment, see equation (11).
Figure BDA0003321448600000084
The square sum of the reactive power Q and the active power P is the square of the apparent power S, and the calculation method is shown in formula (12).
Figure BDA0003321448600000085
Apparent power S equal to voltage effective value UrmsAnd the effective value of the current IrmsThe product of (c) is shown in equation (13).
S＝Urms×Irms (13)
The power factor is usually expressed by λ, and is the ratio of the active power P of the ac circuit to the apparent power S, see equation (14).
λ＝P/S (14)
The collected current data are transformed by combining an equation (15) and a fast Fourier transform algorithm, 32 odd harmonics such as fundamental wave, 3-th harmonic, 5-th harmonic, …, 63-th harmonic and the like are selected as characteristics in the example, the characteristics are respectively marked as Har1, Har3, Har5, … and Har63, and the corresponding harmonic frequencies are respectively 50Hz, 150Hz, 250Hz, … and 3.15 KHz.
Figure BDA0003321448600000091
Where ω is 2 π/T is the angular frequency of the periodic function in equation (15), where ω is 50Hz, T is the period, where 0.02 sec, h is the harmonic order, a0Coefficient of direct current, ah、bhAnd the alternating current coefficient is calculated through Fourier transform.
In order to obtain the trained shannon entropy weighted voting algorithm model in step S8, the method includes the following steps:
step A1: the feature data are stored in a local formed feature database, recursive feature elimination is used for seven algorithms including a linear discriminant analysis algorithm LDA, a naive Bayes classification algorithm NB, a K nearest neighbor algorithm KNN, a decision tree classification algorithm DT, a support vector machine SVM, a logistic regression algorithm LR and a feedback neural network BPNN, recursive feature elimination is used, effective features of the algorithms are selected, the seven algorithms are trained on the basis of the effective features, trained algorithm models are obtained and are recorded as base classifiers, and 7 base classifiers are numbered from 0 to 6.
The Recursive Feature Elimination method (RFE) uses a base model to perform multiple rounds of training, eliminates the features of a plurality of weight coefficients after each round of training, then performs the next round of training based on a new Feature set, and simultaneously performs 10-fold cross validation on the training set in each round of training to obtain the validation accuracy of the current Feature combination. And after the training is finished, finding out the characteristic data combination which has the largest influence on the final prediction result.
FIG. 2 illustrates the effect of the algorithm for different feature quantities of the measured data set on the accuracy of the validation set;
figure 3 illustrates the effect of the algorithm for different feature quantities of the white dataset on the accuracy of the validation set.
The most important characteristics can be screened out through characteristic optimization, and the accuracy of the prediction result is improved.
The seven base classifiers are carried out
Figure BDA0003321448600000092
And combining to obtain 120 different algorithm model combinations.
Step A2: calculating diversity using formula (1) for the 120 algorithm model combinations obtained above;
Figure BDA0003321448600000093
step A3: the average accuracy is calculated by equation (2).
Figure BDA0003321448600000094
To measure ACC in average accuracyiEach basic moldAnd training the model by utilizing the respective optimal characteristic combination, and calculating the accuracy of the test set predicted by the base classifier through a formula (16).
Figure BDA0003321448600000101
Where True (TP) indicates the number of classes for which data carrying Positive class labels are predicted, True Negative (TN) indicates the number of classes for which data carrying Negative class labels are predicted, False Positive (FP) indicates the number of classes for which data carrying Negative class labels are predicted, and False Negative (FN) indicates the number of classes for which data carrying Positive class labels are predicted.
Step A4: calculating the optimal range Ent of the diversity according to the formula (3) by using the maximum value, the minimum value and the midpoint of the diversity*。
Figure BDA0003321448600000102
Step A5: and (4) performing weighted fusion on the prediction classification probability according to the formula (4).
The Shannon entropy weighting voting algorithm is to adopt E classifiers for weighting and fusing the classification problems of n classes, wherein x in the formula (4)kIs a sample, p, in the feature data set XEn(xk) Represents the E-th classifier prediction xkIs the posterior probability of the nth class. w is aiIs the ith classifier on sample xkFusion weight of wiCalculated according to the formula (5), H in the formula (5)i(xk) Shannon entropy, H, representing the ith classifier for all classification classesi(xk) The calculation was performed according to equation (6).
Figure BDA0003321448600000103
Figure BDA0003321448600000104
Figure BDA0003321448600000105
Step A6: adding each column in the formula (4) to realize the posterior probability weighted fusion of the categories to obtain P' (x)k) P' (x) as shown in equation (7)k) The largest column index in the set is for sample xkThe prediction category of (2) is shown in formula (8).
Figure BDA0003321448600000111
Class(xk)＝argmax(P″(xk)) (8)
For the object belonging to Ent*The combination in the range calculates the accuracy of Shannon entropy weighted voting, and the accuracy is recorded as ACCse. Finding the maximum ACCseAnd Ent*And the combination closer to the middle position, namely the combination at the moment is the optimal learner combination which is marked as Com, namely the Shannon entropy weighted voting algorithm model.
The finding process of the optimal learner combination Com is shown in fig. 4, and the combination in the optimal range is found by calculating diversity and determining the optimal range of diversity, which is assumed to include C1. And calculating the Shannon entropy weighted voting accuracy of the C1 learner combinations, and finding the combination Com which is close to the middle position and has the highest Shannon entropy weighted voting accuracy in the optimal range.
For the measured data set used in this example, see FIG. 5, its diversity EntmaxAnd EntminRespectively 0.495 and 0.0078. Viewing scope Ent*Is 0.2514, the range increment σ is set to 0.033. Optimal learner combinations of measured data sets [0,1,3,5]Wherein 0,1,3,5 respectively represent LDA, NB, DT, LR.
When the data set adopts a WHITED data set, see FIG. 6, its diversity EntmaxAnd EntminAre respectively 0.0473And 0.0013, diversity viewing Range Ent*Is 0.0243 and the range increment sigma is set to 0.003. The best combination of the WHITED datasets is [2, 3,5, 6 ]]And represent four base classifiers of KNN, DT, LR and BPNN.
The optimal model combination obtained for the self-testing data set comprises four base classifiers of LDA, NB, DT and LR, and the optimal model combination obtained for WHITED comprises four base classifiers of KNN, DT, LR and BPNN. The features extracted by the base classifier on the two data sets and the best features used for the number of features are shown in table 1.
TABLE 1 features and number of features extracted by the base classifier on two datasets
Figure BDA0003321448600000112
It can be seen that the obtained optimal combination Com of the classification algorithms is different when different data sets are used, but the corresponding optimal combination can be obtained by using the method.
Example 2:
this example uses the non-intrusive load monitoring (NILM) framework of fig. 7 to acquire raw data and uses the self-test data set and white common data set to test the robustness of the algorithm of the present invention.
In the NILM frame, the power frequency of the intelligent electric meter is 50Hz, the sampling frequency is 6.4KHz, so that one period of data can be obtained every 0.02 second, each period comprises 128 current data points and voltage data points, and the data can be divided into high-frequency data and low-frequency data according to the sampling frequency. The low frequency data is data with a frequency lower than 1Hz, and the high frequency data is data higher than 1Hz, so the data used in this example is high frequency data.
According to the multi-thread framework shown in fig. 1, the method employs multi-thread processing, including a main thread and six sub-threads. When multithreading is initiated, data is sent and processed using the signal function and the slot function. The main thread receives original signal data of the instrument through a serial port, analyzes the original signal data to obtain original data, and starts the sub-thread 1 to store the original data in the local in real time.
At the same time, sub-thread 2 is activated, displaying the raw data in the text box of the display device. The original data is transmitted to a protocol analysis module, and then the analyzed voltage and current are transmitted to a feature extraction module to obtain feature data.
Meanwhile, the sub-thread 3 is started to store the obtained real-time characteristic data locally so as to record the characteristics of the equipment in operation.
While, while the main thread is running, the sub-thread 4 is activated, for displaying an image,
the sub-thread 5 is used for capturing the latest feature data in real time and sending the latest feature data to the real-time visual data of the sub-thread 4 for displaying.
In this way, the feature data updated in real time is presented in real time via the display device. In the processing of data by thread 4, the first step is to detect an event and the second step is to start thread 6 and send the latest feature data via signal 6.
And the thread 6 loads a trained algorithm and combines the received latest characteristic data to realize the online identification function of the electric equipment.
The event detection algorithm, which is adopted by the invention, is a Two-sided Cumulative Sum algorithm (CUSUM), and the principle of the algorithm is shown in FIG. 8. W1The window is used to calculate the mean value mu of the sample sequence0；W2The window is the basis for judging whether an event occurs. When W is2Inner fk +And gradually accumulating, and judging that a load event exists when a certain threshold value h is exceeded. If no load event is detected, W2Inner fk +The value is not changed much, at this time W1And W2And sliding the window to the right to a new sampling point, and continuing to detect.
As shown in the formula (16), the algorithm has strong noise interference resistance and high accuracy, and the steady-state event can be accurately identified through the algorithm.
Figure BDA0003321448600000131
Wherein xkIndicates the currently adopted sequence, mu0Is the average of the sequence, usually known or estimable; θ is random noise. Statistic f when sequence is in steady statek +And fk -Fluctuating randomly around the 0 value.
In this example, the real data and the white data set of the real appliances and the combination types thereof in table 3 are collected, and the effect of the method of the present invention is verified under the data of the white data set.
30 electric appliances with steady-state current waveforms in the WHITED are classified and identified, and the robustness of the method provided by the invention is finally verified by comparing an AdaBoost algorithm based on a decision tree, a random forest algorithm and a Stacking algorithm based on a multilayer perceptron (MLP).
For the measured data set, 128 data points are obtained in a single period after the data are analyzed by the detection system, and 38 types of electric appliances are obtained after the data are stored, as shown in table 3.
TABLE 3 actual tested appliances and their mix situation categories
Figure BDA0003321448600000132
Figure BDA0003321448600000141
Each appliance was sampled for 10 seconds to obtain 500 cycles of voltage and current waveform data, each cycle containing 128 sample points. Each numbered sample in table 1 was taken 20 times and the voltage and current data during steady state was averaged 20 times to reduce noise. The measurement is repeated in this way, and a total of 38 × 500 × 128 current data and voltage data of the same magnitude of the above 38 appliance combinations are obtained for training, wherein the data of each period is taken as a training sample.
An additional 1 second of data was collected for testing to ensure that the test data was fresh for the training data, i.e., the test set contained 38 x 50 x 128 current data and the same magnitude of voltage data.
Feature extraction is performed next, and the present invention acquires 36 features in total, which will be described in detail below.
Calculating the active power according to equation (11)
Figure BDA0003321448600000142
As the 33 rd feature.
Calculating reactive power according to equation (12)
Figure BDA0003321448600000151
As the 34 th feature.
Calculating apparent power S ═ U according to equation (13)rms×IrmsAs the 35 th feature.
As the 36 th feature, the power factor λ is calculated according to equation (14).
The amplitudes of the respective odd harmonics are calculated according to equation (15), and there are 32 harmonics in total as the 1 st to 32 nd features.
The method extracts 32 harmonics of apparent power, active power, reactive power, power factors and signals from the acquired current and voltage waveforms of each period as features, converts a training set of original data into vectors of 38 multiplied by 500 multiplied by 36, and each test-numbered electrical appliance comprises 500 samples and each sample comprises 36 features. Similarly, the test set was converted to a 38 × 50 × 36 vector, with 50 samples for each numbered appliance.
All data analyses were calculated using a dell instron 7591 computer, i5-9300HCPU, dominant frequency 2.40 GHz. Python3.7 was used as the programming environment.
The public data set white, still using 50Hz power frequency, has a sampling frequency of 44.1KHz, and contains 882 data points per cycle. By adopting the principle, 30 pieces of electrical appliance data with steady-state current waveforms in the WHITED are obtained and subjected to characteristic conversion. The harmonic features select the first 32 odd harmonics, so as to obtain a training set vector of 30 × 500 × 36. Similarly, a 30 × 50 × 36 test set vector is obtained.
After the characteristics are stored in a locally formed characteristic database, the invention uses recursive characteristic elimination and recursive characteristic elimination method RFE for seven algorithms including a linear discriminant analysis algorithm LDA, a naive Bayes classification algorithm NB, a K nearest neighbor algorithm KNN, a decision tree classification algorithm DT, a support vector machine SVM, a logistic regression algorithm LR and a feedback neural network BPNN, thereby selecting the most effective characteristic data aiming at each algorithm, respectively training the seven algorithms based on the effective characteristics, obtaining the trained algorithm model and obtaining 7 base classifiers with the numbers of 0-6.
Namely a base classifier: LDA 0, NB 1, KNN 2, SVM 3, DT 4, LR 5 and BPNN 6, wherein each base classifier is realized by calling a Sciket-Learn library.
Wherein, the parameters of the LDA model are default.
With NB model, parameters default.
The KNN algorithm is a neighbor algorithm and one of the simplest methods in a classification algorithm, a linear scanning mode is adopted, the k value is set to be 5, and the threshold value of the number of leaf nodes of the tree is set to be 30.
The SVM algorithm adopts a linear kernel function, and the penalty value represents the penalty degree of the error sample and is set to be 1.0.
The DT algorithm is a tree structure that uses Gini as a feature selection criterion, the depth of the tree is 150, and Best as a partition criterion to find the optimal partition point.
With LR model, parameters default.
The BPNN model includes an input layer, two hidden layers and an output layer. For the measured data set, 36 nodes are set on an input layer, 80 nodes are set on two hidden layers, 38 nodes set on an output layer represent 38 electric appliance types which are put into use, Relu is used as an activation function, the learning rate is 0.03, iteration is carried out for 320 times, and Adam is used for optimizing loss. For the white data set, the input layer and the hidden layer are unchanged, and the output layer is set to 30, representing the 30 appliance types selected in the white data set.
The seven models are combined with each other to obtain 120 different model combinations from pairwise combination to seven combinations.
The diversity is calculated using the non-classical entropy formula (1),
Figure BDA0003321448600000161
m is the number of base classifiers in a certain algorithm model combination, and K is the total amount of samples in the characteristic data set X; z (x)k) For a combination of representative algorithmic models, the characteristic data sample xkThe number of the base learners of which the predicted result is consistent with the real result;
calculating the average accuracy by the average accuracy formula (2)
Figure BDA0003321448600000162
m is the number of base classifiers in a certain algorithm model combination;
to measure ACC in average accuracyiAnd training the models by utilizing the respective optimal characteristic combinations of the base models, and calculating the accuracy of the test set through a formula (16).
Figure BDA0003321448600000163
Maximum Ent with diversitymaxMinimum value EntminAccording to formula (3)
Figure BDA0003321448600000164
Calculating the optimal range Ent of the diversity*。
The Shannon entropy weighting voting algorithm is to adopt E classifiers to carry out weighting fusion on n categories of classification problems, as shown in a formula (4)
Figure BDA0003321448600000165
Wherein xkIs a sample in the data set X, pEn(xk) Substitute for Chinese traditional medicineThe posterior probability that the E-th classifier predicts the nth class is shown. Wherein wiIs the ith classifier on sample xkThe fusion weight of (2) is shown in formula (5)
Figure BDA0003321448600000171
Wherein for sample xk，Hi(xk) The calculation method of (2) is shown in formula (6), which represents the shannon entropy of the ith classifier for all classes, and j represents the jth classification class.
Figure BDA0003321448600000172
Adding the columns of the formula (4) to realize the posterior probability weighted fusion of the categories to obtain P' (x)k) As shown in the formula (7),
Figure BDA0003321448600000173
P″(xk) The largest column index in the set is for sample xkThe prediction category of (2), as shown in formula (8)
Class(xk)＝argmax(P″(xk)) (8)。
The finding process of the optimal learner combination Com is shown in FIG. 4, which belongs to Ent*The combination in the range calculates the accuracy of Shannon entropy weighted voting, and the accuracy is recorded as ACCse. Finding the maximum ACCseAnd Ent*The combination closer to the middle position, the combination at this time is the best learner combination, denoted as Com, and the specific algorithm is shown in algorithm 1.
Figure BDA0003321448600000174
Figure BDA0003321448600000181
For the measured data set, see FIG. 5, its diversity EntmaxAnd EntminRespectively 0.495 and 0.0078. Viewing scope Ent*Is 0.2514, the range increment σ is set to 0.033. Optimal learner combinations of measured data sets [0,1,3,5]Wherein 0,1,3,5 respectively represent LDA, NB, DT, LR.
Corresponding to the WHITED dataset, see FIG. 6, its diversity EntmaxAnd Entmin0.0473 and 0.0013, respectively, diversity viewing Range Ent*Is 0.0243 and the range increment sigma is set to 0.003. The best combination of the WHITED datasets is [2, 3,5, 6 ]]And represent four base classifiers of KNN, DT, LR and BPNN.
In the figure, the ordinate is the average accuracy Avg of the learner combination, the abscissa is its diversity Ent, and a dot is a base classifier combination. The shannon entropy weighted voting accuracy of the base classifier combination is calculated at the same time, and is indicated by a star in the figure. The learner combination to be selected should try to maintain Ent*The high average accuracy of the test set in the range is ensured, the high diversity is ensured, and the combined Shannon entropy weighting voting algorithm has the highest accuracy. Therefore, the selection of the optimal combination is generally close to the middle part, and the triangle in the figure represents the combination with the highest accuracy of the shannon entropy weighting voting algorithm, and the diamond corresponding to the combination is the balance point at which the average accuracy and the diversity are balanced, and is also the sought optimal learner combination.
For the self-test dataset, the average accuracy of the four classifiers LDA, NB, DT, LR and the shannon entropy weighting algorithm on the test set was 98.6%, 97.8%, 53.8% and 99.5%, respectively, see fig. 9.
For the white dataset, the average accuracy of the KNN, DT, LR, BPNN four classifiers and the shannon entropy weighting algorithm on the test set was 98.7%, 99.1%, 95.8%, 98.5% and 100%, respectively, see fig. 10.
For both data sets, the classifiers select the best feature combination according to features, and table 1 shows the feature combinations of the four classifiers.
TABLE 2 features and number of features extracted by the base classifier on two datasets
Figure BDA0003321448600000191
The test confusion matrix of the algorithm of the present invention on the measured data set is shown in fig. 11, and it can be seen that 7 samples in the category 26 (desktop + laptop + egg boiler + electric hot wind 2) are predicted to be 19 (desktop + egg boiler + electric hot wind 2). 3 samples in class 27 (desktop + laptop + eggbeater + electric blanket 2) are predicted to be class 16 (desktop + laptop + eggbeater). Looking at the 10 samples tested in the four base classifiers, they were all found to be misclassified and thus still misclassified after integration.
Further comparing the shannon entropy weighted voting algorithm provided herein with other weighting algorithms (see table 4), the result is shown in table 5, which shows that the shannon entropy weighted voting algorithm has higher accuracy, and the robustness of the algorithm is higher than that of other weighting methods.
Table 4 common 4 weights
Figure BDA0003321448600000192
Figure BDA0003321448600000201
TABLE 5 comparison of Algorithm accuracy for different weights
Figure BDA0003321448600000202
Further, the shannon entropy weighting integration algorithm provided by the method is compared with other integration learning algorithms commonly used in the current literature, and mainly comprises an AdaBoost algorithm based on a decision tree, and a Stacking algorithm taking a random forest and a meta model as MLP (modeling and learning) algorithm. The test accuracy of the measured data set and the WHITED public data set is calculated. The parameters and accuracy results for each algorithm used are shown in table 6. It can be seen that the shannon entropy weighted voting-based algorithm provided by the invention has a significant improvement in the test accuracy of both data sets.
TABLE 6 Multi-model accuracy comparison of measured data
Figure BDA0003321448600000203
Figure BDA0003321448600000211
By the above principle, an optimal model combination is obtained for the self-test data set. Next to save, the present invention saves the model using "Joblib" to make the model persistent.
Model reading and prediction, the invention adopts 'Joblib' to load the model in 6 types of sub-threads, and inputs the real-time updated characteristic sequence into the model to predict the classification result.
According to table 7, the time required for five time consuming operations is shown. If a single thread is used, the total time consumed is their sum, i.e. 44 ms.
In the case of multithreading, the total time consumed depends mainly on the main thread, being the sum of the first two threads, i.e. 6 ms. According to table 8, the total time consumed by the four algorithms LDA, NB, DT, LR is small for the measured data. However, the accuracy of the shannon entropy based weighted voting algorithm is highest according to the recognition accuracy.
The industrial frequency in many countries is 50Hz, 0.02S is required for one cycle of data, the algorithm running time required is less than 0.02, and the single thread is definitely timed out since the algorithm time needs less than 20 ms.
TABLE 7 time spent in time consuming operation
Figure BDA0003321448600000212
TABLE 8 time comparison of five algorithms for single thread and multithreading
Figure BDA0003321448600000213
The real-time detection process is shown in fig. 12, after an event is detected, the electric hair drier is obtained through a steady-state data analysis result, and the visual data are obtained in a mode of displaying an active power oscillogram.
The multithreading-based non-intrusive load real-time online monitoring and identifying system provided by the invention based on the Shannon entropy weighted voting algorithm can meet the requirements of real-time tasks and has higher identifying precision.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims (10)

1. The real-time online monitoring and identification method for the power load is characterized in that a Shannon entropy weighted voting algorithm model is used for identifying the type and the working state of the power load, and the determination process of a base classifier combination in the Shannon entropy weighted voting algorithm model is as follows:
selecting E base classifiers to carry out
Figure FDA0003321448590000011
Combining to obtain a plurality of different algorithm model combinations, calculating the diversity Ent of the different algorithm model combinations, and determining the optimal value range Ent of Ent*；
For the object belonging to Ent*The combination of the algorithm models in the range, the accuracy of the Shannon entropy weighted voting of the combination of the algorithm models is calculated and is recorded as ACCse；
Ent*Inner distance Ent*ACC with closest intermediate point and largest numerical valueseAnd the corresponding algorithm model combination is the optimal learner combination which is marked as Com, and the trained Shannon entropy weighted voting algorithm model is obtained.
2. The method for real-time online monitoring and identification of electrical loads according to claim 1, wherein the ACC is used for monitoring and identifying the electrical loadsseThe calculation process is as follows:
1) according to the selected E base classifiers, the prediction classification probability is subjected to weighted fusion according to a formula (4) aiming at the condition of n classification classes
Figure FDA0003321448590000012
P in formula (4)En(xk) Represents the sample x for the feature datakThe E-th base classifier predicts the sample xkIs the posterior probability of the nth class. w is aiIs the ith base classifier to sample xkFusion weight of wiThe calculation formula is as follows:
Figure FDA0003321448590000013
in the formula (5), the reaction mixture is,
Hi(xk) Representing feature-based data samples xkThe Shannon entropy, H, of the ith base classifier calculated for all classification classesi(xk) The calculation formula is as follows:
Figure FDA0003321448590000014
i represents the ith base classifier; j represents the jth class;
2) accumulating the columns in the formula (4) respectively to realize the weighted fusion of the posterior probabilities of the categories to obtain P' (x)k) As shown in formula (7)
Figure FDA0003321448590000021
In the formula (7), the reaction mixture is,
3)P″(xk) The column index of the element with the largest numerical value is the sample x of the feature datakPrediction Class (x) ofk)，
Class(xk)＝arg max(P″(xk)) (8)
arg represents the sequence corresponding to the element with the largest value in the data list; thus, a pair of samples x is obtainedkThe predicted result of (2); according to the steps, obtaining prediction results of all characteristic data samples;
4) calculating the prediction accuracy ACC of each algorithm model combination based on the prediction results and the real results of all training samplesse
Figure FDA0003321448590000022
TPse、TNse、FNse、FPseThe number of the predicted result and the real result of the se combination on the training sample is shown as follows: TPseReally, the number of samples which show real results as positive classes is predicted as the positive classes; TN (twisted nematic)seTrue negative, which represents the number of samples with true results as negative class predicted as negative class; FPseIf the sample is false positive, the number of samples with true results of negative classes is predicted to be positive classes; FN (FN)seIf the true result is false negative, the number of samples with true results of positive class is predicted as negative class; c1 is to select E base classifiers
Figure FDA0003321448590000023
The number of algorithm model combinations obtained by combination.
3. The real-time online electrical load monitoring and identification method according to claim 1, wherein the algorithm model is combined with a diversity Ent and an optimum thereofValue range Ent*The calculation process of (2) is as follows:
1) calculating the diversity Ent of each algorithm model combination by using a formula (1);
Figure FDA0003321448590000024
m is the number of base classifiers in a certain algorithm model combination, and K is the total amount of samples in the characteristic data set X; z (x)k) For a combination of representative algorithmic models, the characteristic data sample xkThe number of the base learners of which the predicted result is consistent with the real result;
2) calculating the optimal range Ent of the diversity according to the formula (3) by using the maximum value and the minimum value of the diversity Ent*
Figure FDA0003321448590000031
In the formula (3), σ is Ent*The half interval range of (a) is an empirical setting, taking a number between 0 and 1.
4. The real-time online monitoring and identification method for the electric load according to any one of claims 1 to 3, characterized by comprising the following steps:
s1, receiving an original signal sent by the intelligent electric meter by using the main thread, and analyzing according to a protocol to obtain original data;
s2, storing the original data by the sub thread 1;
s4, analyzing the original data by the main thread, and extracting characteristic data;
s5, storing the characteristic data to the local in real time by the sub-thread 3 to form a characteristic data set;
s6, the sub-thread 4 calls a display function to initialize the oscillogram of the electric appliance to be displayed;
s7, the sub-thread 5 reads the characteristic data in real time and transmits the characteristic data to the display function in the thread 4 to update the characteristic data in the display;
s8, the sub-thread 4 detects whether an event occurs according to the real-time characteristic data; the event means that the change value of the characteristic data is higher than a set change threshold value in a set detection period;
s9, the sub-thread 6 loads the trained Shannon entropy weighting voting algorithm model, reads real-time characteristic data and inputs the real-time characteristic data into the Shannon entropy weighting voting algorithm model to realize the online identification of the type and the working state of the electric appliance.
5. The real-time online monitoring and identification method for electrical loads according to claim 4, wherein after S2, before S4, there are further steps S3: the sub-thread 2 displays the original data in real time through a display device; the display device displays the raw data in a text format.
6. The real-time online monitoring and identification method for the electrical load according to claim 4, wherein the waveforms in step S6 include voltage waveform, current waveform, active power waveform, reactive power waveform, V-I trace and harmonic amplitude.
7. The real-time online monitoring and identification method for the electrical load according to claim 4, wherein the characteristic data in step S4 includes active power (P), reactive power (Q), apparent power (S), power factor (λ), and (2g-1) subharmonic, where g is a positive integer.
8. The method for real-time online monitoring and identification of electrical loads according to claim 4, wherein in step S5, the feature data are stored locally to form a feature database, the main thread uses recursive feature elimination for a plurality of basic classification algorithms to select the feature data effective for each basic classification algorithm, and the basic classification algorithms are trained based on the corresponding effective feature data to obtain a plurality of trained basic classification algorithms, which are called as basic classifiers.
9. The real-time online power load monitoring and identifying method according to claim 8, wherein the basic classification algorithm comprises at least 2 of a linear discriminant analysis algorithm, a naive bayes classification algorithm, a K-nearest neighbor algorithm, a decision tree classification algorithm, a support vector machine, a logistic regression algorithm, and a feedback neural network.
10. The method as claimed in claim 4, wherein the sub-thread 4 detects the occurrence of an event by using a bilateral accumulation sum algorithm.
CN202111247143.9A 2021-10-26 2021-10-26 Real-time on-line monitoring and identifying method for power load Active CN113947159B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111247143.9A CN113947159B (en) 2021-10-26 2021-10-26 Real-time on-line monitoring and identifying method for power load

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111247143.9A CN113947159B (en) 2021-10-26 2021-10-26 Real-time on-line monitoring and identifying method for power load

Publications (2)

Publication Number Publication Date
CN113947159A true CN113947159A (en) 2022-01-18
CN113947159B CN113947159B (en) 2023-09-05

Family

ID=79332415

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111247143.9A Active CN113947159B (en) 2021-10-26 2021-10-26 Real-time on-line monitoring and identifying method for power load

Country Status (1)

Country Link
CN (1) CN113947159B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117992856A (en) * 2024-04-03 2024-05-07 国网山东省电力公司营销服务中心(计量中心) User electricity behavior analysis method, system, device, medium and program product

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
AU2004202199A1 (en) * 1999-04-30 2004-06-17 E. I. Du Pont De Nemours And Company Distributed hierarchical evolutionary modeling and visualization of empirical data
US20060059112A1 (en) * 2004-08-25 2006-03-16 Jie Cheng Machine learning with robust estimation, bayesian classification and model stacking
US8560134B1 (en) * 2010-09-10 2013-10-15 Kwangduk Douglas Lee System and method for electric load recognition from centrally monitored power signal and its application to home energy management
CN109492667A (en) * 2018-10-08 2019-03-19 国网天津市电力公司电力科学研究院 A kind of feature selecting discrimination method for non-intrusive electrical load monitoring
CN111428755A (en) * 2020-02-29 2020-07-17 国网(苏州)城市能源研究院有限责任公司 Non-invasive load monitoring method
AU2020104000A4 (en) * 2020-12-10 2021-02-18 Guangxi University Short-term Load Forecasting Method Based on TCN and IPSO-LSSVM Combined Model

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
AU2004202199A1 (en) * 1999-04-30 2004-06-17 E. I. Du Pont De Nemours And Company Distributed hierarchical evolutionary modeling and visualization of empirical data
US20060059112A1 (en) * 2004-08-25 2006-03-16 Jie Cheng Machine learning with robust estimation, bayesian classification and model stacking
US8560134B1 (en) * 2010-09-10 2013-10-15 Kwangduk Douglas Lee System and method for electric load recognition from centrally monitored power signal and its application to home energy management
CN109492667A (en) * 2018-10-08 2019-03-19 国网天津市电力公司电力科学研究院 A kind of feature selecting discrimination method for non-intrusive electrical load monitoring
CN111428755A (en) * 2020-02-29 2020-07-17 国网(苏州)城市能源研究院有限责任公司 Non-invasive load monitoring method
AU2020104000A4 (en) * 2020-12-10 2021-02-18 Guangxi University Short-term Load Forecasting Method Based on TCN and IPSO-LSSVM Combined Model

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117992856A (en) * 2024-04-03 2024-05-07 国网山东省电力公司营销服务中心(计量中心) User electricity behavior analysis method, system, device, medium and program product

Also Published As

Publication number Publication date
CN113947159B (en) 2023-09-05

Similar Documents

Publication Publication Date Title
Himeur et al. Robust event-based non-intrusive appliance recognition using multi-scale wavelet packet tree and ensemble bagging tree
Kang et al. Household appliance classification using lower odd-numbered harmonics and the bagging decision tree
US9104189B2 (en) Methods and apparatuses for monitoring energy consumption and related operations
Saini et al. Classification of power quality events–a review
Wang et al. Sample efficient home power anomaly detection in real time using semi-supervised learning
CN109145949A (en) Non-intrusive electrical load monitoring and decomposition method and system based on integrated study
CN111830347B (en) Two-stage non-invasive load monitoring method based on event
Ridi et al. ACS-F2—A new database of appliance consumption signatures
CN111639586B (en) Non-invasive load identification model construction method, load identification method and system
Chahine et al. Electric load disaggregation in smart metering using a novel feature extraction method and supervised classification
Botev et al. Detecting non-technical energy losses through structural periodic patterns in AMI data
US20220120586A1 (en) System and method for identifying appliances under recall
CN109359665A (en) A kind of family's electric load recognition methods and device based on support vector machines
Anderson Non-intrusive load monitoring: Disaggregation of energy by unsupervised power consumption clustering
CN114970633B (en) LSTM-based non-invasive electrical appliance identification method, system and equipment
Bilski et al. Generalized algorithm for the non-intrusive identification of electrical appliances in the household
Rehman et al. Comparative evaluation of machine learning models and input feature space for non-intrusive load monitoring
Park et al. Appliance identification algorithm for a non-intrusive home energy monitor using cogent confabulation
Tongta et al. Long short-term memory (LSTM) neural networks applied to energy disaggregation
CN113947159A (en) Real-time online monitoring and identification method for electrical load
Makonin Approaches to non-intrusive load monitoring (nilm) in the home
CN113193654A (en) Event-driven non-intrusive power load monitoring method based on transient and steady state combination characteristics
Paul et al. Residential appliance identification using 1-D convolutional neural network based on multiscale sinusoidal initializers
Figueiredo Contributions to electrical energy disaggregation in a smart home
Liu et al. Detection of electric bicycle indoor charging for electrical safety: A nilm approach

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant