CN112580471A - Non-invasive load identification method based on AdaBoost feature extraction and RNN model - Google Patents
Non-invasive load identification method based on AdaBoost feature extraction and RNN model Download PDFInfo
- Publication number
- CN112580471A CN112580471A CN202011460508.1A CN202011460508A CN112580471A CN 112580471 A CN112580471 A CN 112580471A CN 202011460508 A CN202011460508 A CN 202011460508A CN 112580471 A CN112580471 A CN 112580471A
- Authority
- CN
- China
- Prior art keywords
- load
- sample
- rnn
- samples
- domain
- 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 48
- 238000000605 extraction Methods 0.000 title claims abstract description 17
- 210000002569 neuron Anatomy 0.000 claims description 27
- 238000012549 training Methods 0.000 claims description 27
- 230000006870 function Effects 0.000 claims description 23
- 230000008859 change Effects 0.000 claims description 12
- 238000012216 screening Methods 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 8
- 230000000694 effects Effects 0.000 claims description 8
- 238000004458 analytical method Methods 0.000 claims description 6
- 150000001875 compounds Chemical class 0.000 claims description 3
- 238000002790 cross-validation Methods 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 230000003213 activating effect Effects 0.000 claims 1
- 238000012544 monitoring process Methods 0.000 abstract description 14
- 238000001514 detection method Methods 0.000 abstract description 6
- 238000013135 deep learning Methods 0.000 abstract description 4
- 230000007246 mechanism Effects 0.000 abstract description 2
- 238000005457 optimization Methods 0.000 abstract description 2
- 230000008569 process Effects 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 2
- 238000013095 identification testing Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000001052 transient effect Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000000691 measurement method Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000035772 mutation Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- 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
- 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
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Molecular Biology (AREA)
- Software Systems (AREA)
- Evolutionary Biology (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computing Systems (AREA)
- Signal Processing (AREA)
- Probability & Statistics with Applications (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a non-invasive load identification method based on AdaBoost feature extraction and deep learning RNN model, which adopts a load detection mechanism based on a self-adaptive variable point optimization algorithm to detect load events. And when a load event is detected, extracting the load characteristics of the electric appliance and storing the load characteristics into the MySQL database to establish a load characteristic library. And designing an RNN (radio network node) suitable for non-invasive load identification, obtaining a load characteristic identification result through the RNN, and calculating the identification accuracy rate aiming at a single electric appliance and the identification accuracy rate when N electric appliances are combined. And finally, changing parameters of the RNN model, comparing load identification accuracy rates under different settings, and realizing optimal network performance by adjusting network parameters. The invention solves the problems that the identification accuracy rate of the household appliances is not high in non-invasive load monitoring, and various appliance equipment combinations and low-power appliances are difficult to correctly identify.
Description
Technical Field
The invention belongs to the field of power consumption detail monitoring of power loads, and relates to a non-invasive load identification method based on AdaBoost feature extraction and an RNN model.
Background
Reasonable energy efficiency management is beneficial to reducing the electric energy consumption and ensuring the safe operation of the power grid. With the rapid increase of the electricity consumption of residents and the advanced innovation of power systems in recent years, the load monitoring technology becomes a hot spot concerned by people. Load monitoring techniques include traditional intrusive load monitoring and more advanced non-intrusive load monitoring techniques. The intrusive load monitoring is usually to install an intelligent sensor at each load component or each appliance access point to monitor the working state of the load component or the appliance access point. The invasive measurement method has high cost and great maintenance difficulty. In contrast, Non-invasive power monitoring (NILM) is a new method that is very user friendly and economically convenient. The method only needs to install monitoring equipment at the power distribution inlet wire, and analyzes the power utilization information of a user and monitors the load type through a software program. The non-invasive power monitoring is a necessary trend in the era of intelligent electric meters, and not only can provide detailed power utilization information for users, but also can accurately predict load composition proportion for power companies, so that the research on the aspect has important significance.
Many researchers have applied machine learning algorithms to the field of non-intrusive load monitoring technology. However, the existing methods still have the problem of low recognition accuracy for the combination of multiple electrical devices and electrical devices with small power (such as computers and televisions), which seriously affects the implementation of non-intrusive load monitoring and the implementation of intelligent power utilization.
Therefore, it is desirable to have a deep learning non-intrusive load identification method that can achieve higher identification accuracy.
Disclosure of Invention
The invention aims to provide a non-invasive load identification method based on AdaBoost feature extraction and an RNN model aiming at the defects of the prior art, so as to solve the problems that the identification accuracy rate of household appliances is not high, and various appliance equipment combinations and small-power appliances are difficult to correctly identify in non-invasive load monitoring.
The purpose of the invention is realized by the following technical scheme: a non-invasive load identification method based on AdaBoost feature extraction and RNN model comprises the following steps:
the method comprises the following steps: judging a switching event according to the power change at the household power distribution inlet wire, and when the switching event occurs, acquiring an electric appliance current sample of the switching event at the household power distribution inlet wire to form a data set used for follow-up Adaboost characteristic extraction and RNN model training;
step two: carrying out frequency domain analysis on the collected current sample, extracting the frequency domain harmonic amplitude of the current sample as the load characteristics of each electric appliance, wherein the load characteristics of each type of electric appliance form a characteristic domain, and all the characteristic domains form a load characteristic library which comprises the load characteristics of each type of electric appliance;
step three: and simplifying the load feature library obtained in the second step, screening and reserving domain boundary samples of each type of electric appliance feature domain in the feature library by using an AdaBoost algorithm, and reducing the samples in the domain.
Step four: designing an RNN (radio network node) suitable for non-invasive load identification, determining the number of neurons of an input layer and an output layer of the RNN, the number of layers of a hidden layer of the RNN, the number of the neurons and an activation function of each layer of the RNN, and selecting a loss function for the RNN; the number of input layer neurons is determined by the dimension of the input load feature vector, while the dimension of the corresponding expected load tag vector determines the number of output layer neurons. The number of the hidden layers and the number of the neurons are selected according to the actual engineering requirements.
Step five: RNN network parameters are initialized. Wherein the parameters include: weight ω between input layer-hidden layerijWeight ω between hidden layer-output layerjkLearning rate, maximum number of iterations, and error limit (error-gate);
step six: the method comprises the steps of taking various electrical appliance load characteristics screened by an AdaBoost algorithm as input of an RNN model, taking electrical appliance categories as output of the RNN model, training an RNN network by using an acquired training data set, carrying out frequency domain analysis on electrical appliance current samples to be identified, wherein switching events occur, extracting load characteristics, screening by the AdaBoost algorithm, inputting the electrical appliance current samples to the trained RNN model, and obtaining a load characteristic identification result.
Furthermore, a certain time point is used as a boundary to be divided into two types in a time window with a certain length, the time point is classified into one type before and another type after, so that the sum of the variances in the same type is minimum, and the sum of the variances between the types is maximum. At this time, if the distance between the two types of mean values exceeds a limited threshold value, the moment is considered as a change point, and a switching event exists. The defined threshold is determined by the load with the least power.
Further, in the third step, the process of determining whether the sample belongs to the domain boundary sample specifically includes: AdaBoost algorithm proceedsWhen classifying, inputting a sample, outputting the probability P that the sample belongs to the A-type domain, the B-type domain, … … and the N-type domainA,PB,……,PN. When the probability that the input sample belongs to the i-class domain and the j-class domain satisfies the condition: i Pi-PjIf the | is less than 0.1, the input sample is a domain boundary sample of the i-class domain and the j-class domain, and misjudgment is easy to occur, the sample is reserved; and randomly deleting 50% of samples which do not meet the conditions; the specific steps for reducing samples in the domain are as follows:
(1) selecting a sample set X of two types of electric appliances i and j from a load characteristic libraryi,XjAnd composing a sample sequence:
S＝(Xi,Xj)
for the channel formed by miIndividual electric appliances i and mjA sequence S of samples of the individual appliance j, wherein the expected output of each sample is the class i or j of the sample, so the expected output S of the sequence S of samples is:
(2) initialization: setting the total number of the two types of load samples as m, initializing the weight of the load sample sequence:
D1(u)＝1/m
wherein m is mi+mj,u＝1,2,……,m
(3) And (3) weak classifier classification: for the t-th weak classifier, the classification error is calculated according to the weight of the classification error sample:
wherein l is a classification error sample, gt(l) For the classification result of the t-th weak classifier on the classification error sample, s (l) is the expected classification result on the classification error sample.
(4) T weak classifier weight:
(5) adjusting sample weight:
wherein S (u) is the expected classification result for all load samples, gt(u) the result of the classification of all the load samples by the t-th classifier, BtIs a normalization factor.
(6) And (5) repeating the steps (2) to (5) until the weak classifiers with the number n set according to the requirements are trained, and obtaining the final weight of the electric appliance sample. The number n of weak classifiers is generally 50-200. If a better effect is desired, cross-validation can be used for parameter adjustment.
(7) And changing the type of the sample, and repeating the steps to complete the weight calculation between every two N electric appliances. After the weight calculation is finished, for the electric appliance combination with the changed weight, selecting a sample with the weight more than 0.008 from the electric appliance combination, and reserving the sample, wherein the sample with the weight less than 0.002 and the sample with the unchanged weight is randomly deleted by 50% of the sample amount.
Further, the number of input layer neurons is 2NAnd the number of neurons in the output layer is K. The output variable represents the state of the electric appliance output and stop with 0, and 1 represents the state of the electric appliance output and running.
Further, the activation function of the RNN network in step four is:
in the formula, the parameter is usually set to λ 1 and b 0.
Further, the loss function of the RNN network in step four is a cross-entropy loss function, as follows:
in the formula (I), the compound is shown in the specification,
to output the desired output value of the layer, ytFor the true value, T represents the number of learning samples, and N represents the dimension of the output quantity.
Further, the specific operation of initializing RNN network parameters is as follows. OmegaijInitialisation to a random number in the range (-1,1), omegajkInitialization is 0, learning rate is initialized to 0.01, error limit error _ gate is set to 1 × 10-4The maximum number of iterations was set to 10000.
Further, in the sixth step, after the load characteristic identification result is obtained, the identification accuracy rate for a single electric appliance and the identification accuracy rate when the N electric appliances are combined are calculated; and then changing the number of hidden layer neurons, the number of hidden layer layers, the maximum iteration number and the value of error-gate of the RNN model, comparing the load identification accuracy and the learning curve under different parameter settings, and realizing the optimal network performance by adjusting network parameters.
Further, in the sixth step, the identification accuracy for a single electrical appliance and the identification accuracy when N electrical appliances are combined are calculated as follows:
in the formula, MiFor the total number of samples of each type, i is 1,2, … …, N, miThe number correctly identified in each type of appliance sample.
The invention has the advantages and effects that:
1. the technical method can realize the estimation of the power consumption of each target electrical appliance in the load only by analyzing the total data of the power load at the household power inlet, namely realize the non-invasive load monitoring and identification. The method can obviously improve the accuracy of load identification, and can obtain the effect exceeding that of the prior art on the load identification of complex electrical equipment and low-power electrical equipment.
2. The technical method of the invention uses AdaBoost algorithm to carry out feature screening and simplification on the extracted load features, reserves domain boundary samples of each class in the feature space, and reduces the samples in the domain. The method achieves the aims of reducing the calculated amount and improving the load identification efficiency, and realizes the real-time online identification of the load.
3. The technical method of the invention applies a deep learning model-RNN to solve the problem of non-invasive power load identification, and establishes the internal association of input mapping to output by the characteristic that the RNN has the memory history input characteristic quantity and the like, thereby realizing the load identification of time series input. In order to further improve the load identification capability, a large number of experiments are carried out, and the optimal activation function, the loss function and the RNN model parameters are selected, so that a satisfactory load identification effect is obtained. The invention adopts the deep learning technology, and can learn the complex mode which is difficult to be processed and distinguished manually, so the technology of the invention has better market application prospect and large-range popularization value.
4. The invention adopts an end-to-end technical scheme. When the load identification is completed, the required various load characteristics do not need to be selected and extracted manually, and only the total data of the power utilization parameters need to be input into the algorithm model. The end-to-end scheme ensures that the invention has higher user-friendliness and usability and can be widely applied to identification of the electrical load of residents.
Drawings
Fig. 1 is an overall architecture diagram of a non-invasive load identification method based on AdaBoost load feature extraction and an RNN model.
Fig. 2 is a schematic distribution diagram of two electrical samples with similar characteristics in a characteristic space.
Fig. 3 is a graph showing a change of an effective current value within a certain period of time when the load event is detected according to the embodiment of the present invention.
Fig. 4 is a training framework of the RNN network of the present invention.
FIG. 5 shows the variation of the on-line recognition accuracy when different RNN hidden layer node numbers are set.
Detailed Description
The present invention will be further described with reference to the following specific examples, but the present invention is not limited to these examples.
As shown in fig. 1, the present invention provides a non-invasive load identification method based on AdaBoost feature extraction and RNN model, comprising the following steps:
the method comprises the following steps: judging a switching event according to the power change at the household power distribution inlet wire, and when the switching event occurs, acquiring an electric appliance current sample of the switching event at the household power distribution inlet wire to form a data set used for follow-up Adaboost characteristic extraction and RNN model training; the detection method of the switching event adopts a load detection mechanism based on a self-adaptive variable point optimizing algorithm, a certain time point is used as a boundary to be divided into two types in a time window with a certain length, the time point is classified into one type before and another type after, so that the sum of variances in the same type is minimum, and the sum of variances between the types is maximum. At this time, if the distance between the two types of mean values exceeds a limited threshold value, the moment is considered as a change point, and a switching event exists.
The load detection method based on the adaptive variable point optimization algorithm is described in detail as follows. At a certain time point k, the L data samples { x ] in the time windowiAre divided into two categories, then C0Class { x1,x2,…,xkAnd C1Class { xk+1,xk+2,…,xLThe probabilities are respectively:
the mean values are respectively:
in the formula, piIs xiIn the data set { xiThe probability of occurrence in } satisfies
t represents time. According to the classification rule, a certain sample can be calculated to belong to C0Class and C1The probabilities of the classes are:
therefore, when the objective function is satisfied to the minimum, a potential change point can be obtained as shown in the following formula. In combination with the difference between the two types of means, a load event is considered to exist when the difference exceeds some defined threshold, which is also referred to as a change point. Generally, this defined threshold is determined by the load at which power is minimal.
Step two: and carrying out frequency domain analysis on the collected current samples, extracting the frequency domain harmonic amplitude as the load characteristic of each electric appliance, and extracting by taking the power waveform in the switching transient process of the electric appliances as the transient characteristic. And performing Fourier series expansion on the current of the electric appliance in steady-state working, and taking the harmonic amplitude of the current as steady-state characteristics. The load characteristics of each type of electric appliance form a characteristic field, all the characteristic fields form a load characteristic library, the characteristic library comprises the load characteristics of each type of electric appliance, and the obtained load characteristics are stored in a MySQL database.
Step three: and simplifying the load feature library obtained in the second step, screening and reserving domain boundary samples of each type of electric appliance feature domain in the feature library by using an AdaBoost algorithm, and reducing the samples in the domain. As shown in fig. 2, a schematic distribution diagram of two electrical appliance samples with similar characteristics in a characteristic space is shown, and a process of determining whether a sample belongs to a domain boundary sample specifically includes: when the AdaBoost algorithm is used for classification, a sample is input, and the probability P that the sample belongs to the A-type domain, the B-type domain, … … and the N-type domain is outputA,PB,……,PN. When the probability that the input sample belongs to the i-class domain and the j-class domain satisfies the condition: i Pi-PjIf the | is less than 0.1, the input sample is a domain boundary sample of the i-class domain and the j-class domain, and misjudgment is easy to occur, the sample is reserved; and randomly deleting 50% of samples which do not meet the conditions; the specific steps for reducing samples in the domain are as follows:
(1) selecting a sample set X of two types of electric appliances i and j from a load characteristic libraryi,XjAnd composing a sample sequence:
S＝(Xi,Xj)
for the channel formed by miIndividual electric appliances i and mjA sequence S of samples of the individual appliance j, wherein the expected output of each sample is the class i or j of the sample, so the expected output S of the sequence S of samples is:
(2) initialization: setting the total number of the two types of load samples as m, initializing the weight of the load sample sequence:
D1(u)＝1/m
wherein m is mi+mj,u＝1,2,……,m
(3) And (3) weak classifier classification: for the t-th weak classifier, the classification error is calculated according to the weight of the classification error sample:
wherein l is a classification error sample, gt(l) For the classification result of the t-th weak classifier on the classification error sample, s (l) is the expected classification result on the classification error sample.
(4) T weak classifier weight:
(5) adjusting sample weight:
wherein S (u) is the expected classification result for all load samples, gt(u) the result of the classification of all the load samples by the t-th classifier, BtIs a normalization factor.
(6) And (5) repeating the steps (2) to (5) until the weak classifiers with the number n set according to the requirements are trained, and obtaining the final weight of the electric appliance sample. The number n of weak classifiers is generally 50-200. If a better effect is desired, cross-validation can be used for parameter adjustment.
(7) And changing the type of the sample, and repeating the steps to complete the weight calculation between every two N electric appliances. After the weight calculation is finished, for the electric appliance combination with the changed weight, selecting a sample with the weight more than 0.008 from the electric appliance combination, and reserving the sample, wherein the sample with the weight less than 0.002 and the sample with the unchanged weight is randomly deleted by 50% of the sample amount. Compared with the previous sample, the screened sample is greatly simplified, so that the load feature library is simpler, and the calculation amount of the load identification algorithm is effectively reduced under the condition of not influencing the identification accuracy.
Step four: design ofThe method comprises the steps that the number of neurons of an input layer and an output layer of the RNN, the number of layers of a hidden layer of the RNN and the number of the neurons, and an activation function of each layer of the RNN are determined, and a loss function is selected for the RNN; the number of input layer neurons is determined by the dimension of the input load feature vector, while the dimension of the corresponding expected load tag vector determines the number of output layer neurons. The selection of the number of hidden layers and the number of the neurons is generally related to engineering practice, generally speaking, the smaller the number of the neurons is, the poorer the recognition effect under the aliasing characteristics is, and along with the increase of the number of the neurons, the recognition effect is gradually improved firstly and then tends to be stable and not to be improved any more. According to engineering practice experience, input layer selection 2N(N ═ 1,2,3, … …) neurons. The number of neurons in the output layer is K. The output variable represents the state of the electric appliance output and stop with 0, and 1 represents the state of the electric appliance output and running. The input timing variable is { x1,x2,......,xT-wherein the input time sequence of the neuron at the sampling instant t is:
in order to solve the unstable convergence of the network model constructed by the binary function and the linear function, a standard S-shaped activation function is used. The RNN network activation function is:
in the formula, the parameter is usually set to λ 1 and b 0.
In order to solve the problem of gradient disappearance in the RNN training process, the method selects a cross entropy loss function which can minimize the deviation of the output of a training sample set and an expected value, and the loss function of the RNN network is as follows:
in the formula (I), the compound is shown in the specification,
to output the desired output value of the layer, ytFor the true value, T represents the number of learning samples, and N represents the dimension of the output quantity.
Step five: RNN network parameters are initialized. Wherein the parameters include: weight ω between input layer-hidden layerijWeight ω between hidden layer-output layerjkLearning rate, maximum number of iterations, and error limit (error-gate); omegaijInitialisation to a random number in the range (-1,1), omegajkInitialization is 0, learning rate is initialized to 0.01, error limit error _ gate is set to 1 × 10-4The maximum number of iterations was set to 10000.
Step six: various electrical appliance load characteristics screened by the AdaBoost algorithm are used as input of the RNN model, the electrical appliance category is used as output of the RNN model, and the acquired training data set is used for training the RNN network, as shown in FIG. 4, the RNN network training framework is provided in the invention, and the RNN network training method is specifically described as follows: and training the designed network by using the training set samples, and stopping training when the change of the overall recognition accuracy rate is less than the threshold of error _ gate or the maximum iteration number is reached. And (4) carrying out identification test on the test set sample by using the trained RNN, and checking the identification accuracy and the overall identification accuracy of the network to each electric appliance.
And carrying out frequency domain analysis on the electric appliance current sample to be identified, which has a switching event, extracting load characteristics, screening by an AdaBoost algorithm, and inputting the electric appliance current sample to the trained RNN model to obtain a load characteristic identification result. Then calculating the identification accuracy rate aiming at the single electric appliance and the identification accuracy rate when the N electric appliances are combined; the method comprises the following specific steps:
in the formula, MiFor the total number of samples of each type, i is 1,2, … …, N, miFor each type of appliance sampleThe number of correctly identified.
The relationship between the network parameters and the network performance is the training time of the network under different parameters, the identification time of the network to a load event and the overall identification accuracy. In order to achieve better performance and higher working efficiency, the method further selects appropriate network parameter settings through experiments, and specifically comprises the following steps: changing the number of hidden layer neurons, the number of hidden layer layers, the maximum iteration times and the value of error-gate of the RNN model, comparing the load identification accuracy and the learning curve under different parameter settings, and realizing the optimal network performance by adjusting network parameters.
One embodiment of the present invention is as follows:
an overall architecture diagram of a non-invasive load identification method based on AdaBoost load feature extraction and an RNN model is shown in fig. 1. An embodiment and specific implementation steps of the invention are given as follows:
step 1: the invention takes voltage and current signals collected by a certain research institute at the power transmission and distribution main inlet as experimental data. Fig. 3 shows the change of the effective current value in a certain period of time, which includes the switching of the electric equipment. It can be observed that the actual load event occurs at 14 change points, 20, 38, 48, 78, 100, 140, 170, 210, 232, 260, 290, 310, 350, and 380. Table 1 shows the detection results obtained with the time window detection method under different time windows. As can be seen from table 1, the load event occurrence point can be better detected in a shorter time window. The window size is chosen to be 10 according to a number of practical experience.
TABLE 1
Step 2: and when the mutation point is detected in the step 1, indicating that the switching-in or switching-off operation of the electrical equipment occurs. At the moment, the load characteristics are extracted, and the extracted load characteristics are stored in the MySQL database to form a load characteristic database.
And step 3: and (3) explaining an AdaBoost sample screening algorithm by taking the steady-state current harmonic amplitude extracted in the step (2) as a load characteristic. And (3) extracting 42000 groups of current samples of 7 types of electric appliances in steady-state working from the load characteristic database obtained in the step (2), and randomly selecting 21000 groups of samples from the current samples for algorithm verification. And performing Fourier series expansion on 21000 groups of current samples to extract odd harmonic current amplitude as load characteristics, randomly selecting 14000 groups of data as training set samples from the obtained load characteristics of 7 types of electrical appliances, and using the rest 7000 groups of data as test set samples.
14000 training set data were screened using the AdaBoost algorithm. According to the obtained sample weight D between every two 7 types of loads12，D13，……，D67Can find D14，D26And D27The weights of the electric appliances 1 and 4, the electric appliances 2 and 6 and the electric appliances 2 and 7 are almost unchanged, so that misjudgment is easy to occur in training (samples with large weights indicate that misjudgment is easy to occur, and samples with small weights do not easily cause misjudgment). Therefore, the load characteristic samples of the electrical appliances 1,2, 4, 6 and 7 are screened according to the weights thereof, and the load characteristic sample with larger weight is reserved in each electrical appliance. The samples with smaller weight in the 5 kinds of appliances and the samples with other 2 kinds of appliances with almost unchanged weight are not easy to be misjudged because of being in the class domain or being far away from other class domains, and therefore only a small part of the samples are reserved. The method retains 1/4 for these samples, thus forming a new training set from the screened samples.
After screening, the number of the training set samples is changed from 14000 groups to 4000 groups, the training set data is screened through an AdaBoost algorithm, samples which are important for load identification are reserved, and redundant data are removed.
And 4, step 4: and (3) taking 4000 groups of samples screened by the AdaBoost algorithm in the step 3 as a training data set of the RNN. Initializing parameters of each layer of the RNN: the input dimension is 16, the number of hidden layer nodes is 32, and the number of output layer nodes is 7. And adopting a softsign function as an activation function of each layer of the RNN, and adopting a cross entropy error as a loss function.
And 5: step 4, initializing the learning step length of RNN network as0.01, all elements of the weight parameter matrix are initialized to a value within the range of (-1,1), and the error threshold error _ gate is set to 1 × 10-4The maximum number of iterations was set to 10000.
Step 6: the RNN training process is based on the variance between the actual and expected values of the network output, as shown in the following equation:
when the maximum number of iterations is 10000, the average value of errors finally reaches 2.2377 multiplied by 10-4And the training process is finished.
And carrying out online identification test by using the trained RNN model. Firstly, load data is recorded into a MySQL database, wherein sampling time points take 2s as an interval, and the power utilization equipment with 1 in each equipment state combination is started at 1min intervals in sequence. And when the electric equipment with the state of 1 in the equipment state combination is fully opened for more than 20s, sequentially closing the electric equipment. The data of 10 time points under the condition that the electric equipment is fully opened are obtained from the data and the load identification is carried out on the data.
As shown in table 2, the present invention is a summary table of the accuracy of the trained RNN network for performing online load identification on different electrical equipment combinations. As can be seen from the data in the table, when load events occur in any 1 device, any 2 devices, and any 7 devices, the recognition accuracy reaches 100%, and even in any 5 device combinations, the recognition accuracy reaches 83.33%. When the number of the electric equipment combinations is small, the accuracy is high, and when the number of the electric equipment combinations is large, some identification errors may occur. The main reason is the similarity of the superposed harmonic features of the electric equipment, so that aliasing is generated on the features, and the RNN model identification error is caused.
TABLE 2
And 7: in order to obtain the RNN model with the optimal identification performance, load identification is further carried out by selecting different hidden layer node numbers. Fig. 5 shows the load recognition results when the number of hidden layer nodes is 32, 64, and 96, respectively. It can be found that the load identification accuracy is improved to some extent initially when the number of hidden layer nodes increases, but the accuracy is hardly improved any more when the number of nodes increases to some extent. Therefore, in practical application, the number of hidden layer nodes needs to be properly adjusted according to practical situations.
The invention can realize non-invasive high-precision identification on resident power loads and solve the problem that the load identification cannot be accurately carried out under the conditions of complex electrical appliances and various electrical appliance combinations at present. In addition, the technology used by the invention has good real-time performance, and realizes the online identification of the load. By properly adjusting the network parameters, the performance of the model can be further optimized, so that the method can be suitable for each specific scene of residential power load identification.
Finally, it should be pointed out that: the above examples are only for illustrating the technical solutions of the present invention, and are not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.
Claims (9)
1. A non-invasive load identification method based on AdaBoost feature extraction and RNN model is characterized by comprising the following steps:
the method comprises the following steps: judging a switching event according to the power change at the household power distribution inlet wire, and when the switching event occurs, acquiring an electric appliance current sample of the switching event at the household power distribution inlet wire to form a data set used for follow-up Adaboost characteristic extraction and RNN model training;
step two: carrying out frequency domain analysis on the collected current sample, extracting the frequency domain harmonic amplitude of the current sample as the load characteristics of each electric appliance, wherein the load characteristics of each type of electric appliance form a characteristic domain, and all the characteristic domains form a load characteristic library which comprises the load characteristics of each type of electric appliance;
step three: and simplifying the load feature library obtained in the second step, screening and reserving domain boundary samples of each type of electric appliance feature domain in the feature library by using an AdaBoost algorithm, and reducing the samples in the domain.
Step four: designing an RNN (radio network node) suitable for non-invasive load identification, determining the number of neurons of an input layer and an output layer of the RNN, the number of layers of a hidden layer of the RNN, the number of the neurons and an activation function of each layer of the RNN, and selecting a loss function for the RNN; the number of input layer neurons is determined by the dimension of the input load feature vector, while the dimension of the corresponding expected load tag vector determines the number of output layer neurons. The number of the hidden layers and the number of the neurons are selected according to the actual engineering requirements.
Step five: RNN network parameters are initialized. Wherein the parameters include: weight ω between input layer-hidden layerijWeight ω between hidden layer-output layerjkLearning rate, maximum number of iterations, and error limit (error-gate);
step six: the method comprises the steps of taking various electrical appliance load characteristics screened by an AdaBoost algorithm as input of an RNN model, taking electrical appliance categories as output of the RNN model, training an RNN network by using an acquired training data set, carrying out frequency domain analysis on electrical appliance current samples to be identified, wherein switching events occur, extracting load characteristics, screening by the AdaBoost algorithm, inputting the electrical appliance current samples to the trained RNN model, and obtaining a load characteristic identification result.
2. The method as claimed in claim 1, wherein the method is divided into two classes within a time window of a certain length by using a time point as a boundary, the time point is classified into one class before the time point and classified into another class after the time point, so that the intra-class variance sum is minimum, and the inter-class variance sum is maximum. At this time, if the distance between the two types of mean values exceeds a limited threshold value, the moment is considered as a change point, and a switching event exists. The defined threshold is determined by the load with the least power.
3. The method as claimed in claim 1, wherein the step three of determining whether the sample belongs to the domain boundary sample specifically comprises: when the AdaBoost algorithm is used for classification, a sample is input, and the probability P that the sample belongs to the A-type domain, the B-type domain, … … and the N-type domain is outputA,PB,……,PN. When the probability that the input sample belongs to the i-class domain and the j-class domain satisfies the condition: i Pi-PjIf the | is less than 0.1, the input sample is a domain boundary sample of the i-class domain and the j-class domain, and misjudgment is easy to occur, the sample is reserved; and randomly deleting 50% of samples which do not meet the conditions; the specific steps for reducing samples in the domain are as follows:
(1) selecting a sample set X of two types of electric appliances i and j from a load characteristic libraryi,XjAnd composing a sample sequence:
S＝(Xi,Xj)
for the channel formed by miIndividual electric appliances i and mjA sequence S of samples of the individual appliance j, wherein the expected output of each sample is the class i or j of the sample, so the expected output S of the sequence S of samples is:
(2) initialization: setting the total number of the two types of load samples as m, initializing the weight of the load sample sequence:
D1(u)＝1/m
wherein m is mi+mj,u＝1,2,……,m
(3) And (3) weak classifier classification: for the t-th weak classifier, the classification error is calculated according to the weight of the classification error sample:
wherein l is a classification error sample, gt(l) For the classification result of the t-th weak classifier on the classification error sample, s (l) is the expected classification result on the classification error sample.
(4) T weak classifier weight:
(5) adjusting sample weight:
wherein S (u) is the expected classification result for all load samples, gt(u) the result of the classification of all the load samples by the t-th classifier, BtIs a normalization factor.
(6) And (5) repeating the steps (2) to (5) until the weak classifiers with the number n set according to the requirements are trained, and obtaining the final weight of the electric appliance sample. The number n of weak classifiers is generally 50-200. If a better effect is desired, cross-validation can be used for parameter adjustment.
(7) And changing the type of the sample, and repeating the steps to complete the weight calculation between every two N electric appliances. After the weight calculation is finished, for the electric appliance combination with the changed weight, selecting a sample with the weight more than 0.008 from the electric appliance combination, and reserving the sample, wherein the sample with the weight less than 0.002 and the sample with the unchanged weight is randomly deleted by 50% of the sample amount.
4. The method of claim 1, wherein the number of neurons in the input layer is 2NAnd the number of neurons in the output layer is K. Output ofThe variable represents the out and stop state of the appliance with 0, and the 1 represents the out and running state of the appliance.
5. The method of claim 1, wherein the activating function of the RNN network in the fourth step is:
in the formula, the parameter is usually set to λ 1 and b 0.
6. The method for non-invasive load identification based on AdaBoost feature extraction and RNN model according to claim 1, wherein the loss function of RNN in step four is a cross entropy loss function as follows:
in the formula (I), the compound is shown in the specification,
to output the desired output value of the layer, ytFor the true value, T represents the number of learning samples, and N represents the dimension of the output quantity.
7. The method for non-invasive load identification based on AdaBoost feature extraction and RNN model according to claim 1, wherein initializing RNN network parameters is as follows. OmegaijInitialisation to a random number in the range (-1,1), omegajkInitialization is 0, learning rate is initialized to 0.01, error limit error _ gate is set to 1 × 10-4The maximum number of iterations was set to 10000.
8. The non-invasive load identification method based on AdaBoost feature extraction and RNN model according to claim 1, characterized in that in step six, after obtaining the load feature identification result, the identification accuracy rate for a single electric appliance and the identification accuracy rate when N electric appliances are combined are calculated; and then changing the number of hidden layer neurons, the number of hidden layer layers, the maximum iteration number and the value of error-gate of the RNN model, comparing the load identification accuracy and the learning curve under different parameter settings, and realizing the optimal network performance by adjusting network parameters.
9. The method for non-invasive load identification based on AdaBoost feature extraction and RNN model according to claim 8, wherein in step six, the calculation of the identification accuracy for a single electrical appliance and the identification accuracy for the combination of N electrical appliances is specifically as follows:
in the formula, MiFor the total number of samples of each type, i is 1,2, … …, N, miThe number correctly identified in each type of appliance sample.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011460508.1A CN112580471A (en) | 2020-12-11 | 2020-12-11 | Non-invasive load identification method based on AdaBoost feature extraction and RNN model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011460508.1A CN112580471A (en) | 2020-12-11 | 2020-12-11 | Non-invasive load identification method based on AdaBoost feature extraction and RNN model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112580471A true CN112580471A (en) | 2021-03-30 |
Family
ID=75131552
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011460508.1A Pending CN112580471A (en) | 2020-12-11 | 2020-12-11 | Non-invasive load identification method based on AdaBoost feature extraction and RNN model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112580471A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113094931A (en) * | 2021-04-29 | 2021-07-09 | 杭州电力设备制造有限公司 | Non-invasive load decomposition method, device and equipment |
| CN113191253A (en) * | 2021-04-28 | 2021-07-30 | 浙江大学 | Non-invasive load identification method based on feature fusion under edge machine learning |
| CN113223287A (en) * | 2021-04-02 | 2021-08-06 | 国电南瑞南京控制系统有限公司 | SPI-based communication method and system for intelligent electric energy meter and load identification module |
| CN113255236A (en) * | 2021-07-07 | 2021-08-13 | 浙江大学 | Non-invasive load self-adaptive identification method based on twin network |
| CN113361831A (en) * | 2021-08-10 | 2021-09-07 | 国网江苏省电力有限公司营销服务中心 | Non-invasive load identification electric quantity decomposition method and system based on priority distribution |
| CN114019281A (en) * | 2021-11-04 | 2022-02-08 | 国网四川省电力公司营销服务中心 | Non-invasive load monitoring method and system based on die body excavation and semi-supervision method |
| CN114184870A (en) * | 2021-12-14 | 2022-03-15 | 河北科技大学 | Non-invasive load identification method and equipment |
| CN114970633A (en) * | 2022-06-07 | 2022-08-30 | 西安电子科技大学 | LSTM-based non-invasive electrical appliance identification method, system and equipment |
| CN117633456A (en) * | 2023-11-17 | 2024-03-01 | 国网江苏省电力有限公司 | Marine wind power weather event identification method and device based on self-adaptive focus loss |
-
2020
- 2020-12-11 CN CN202011460508.1A patent/CN112580471A/en active Pending
Non-Patent Citations (2)
| Title |
|---|
| 刘恒勇 等: "《一种关联RNN模型的非侵入式负荷辨识方法》", 《电力系统保护与控制》 * |
| 宋旭帆 等: "《基于K-NN结合核Fisher判别的非侵入式负荷监测方法》", 《电力系统自动化》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113223287A (en) * | 2021-04-02 | 2021-08-06 | 国电南瑞南京控制系统有限公司 | SPI-based communication method and system for intelligent electric energy meter and load identification module |
| CN113191253A (en) * | 2021-04-28 | 2021-07-30 | 浙江大学 | Non-invasive load identification method based on feature fusion under edge machine learning |
| CN113094931A (en) * | 2021-04-29 | 2021-07-09 | 杭州电力设备制造有限公司 | Non-invasive load decomposition method, device and equipment |
| CN113094931B (en) * | 2021-04-29 | 2022-04-22 | 杭州电力设备制造有限公司 | Non-invasive load decomposition method, device and equipment |
| CN113255236A (en) * | 2021-07-07 | 2021-08-13 | 浙江大学 | Non-invasive load self-adaptive identification method based on twin network |
| CN113361831A (en) * | 2021-08-10 | 2021-09-07 | 国网江苏省电力有限公司营销服务中心 | Non-invasive load identification electric quantity decomposition method and system based on priority distribution |
| CN114019281A (en) * | 2021-11-04 | 2022-02-08 | 国网四川省电力公司营销服务中心 | Non-invasive load monitoring method and system based on die body excavation and semi-supervision method |
| CN114019281B (en) * | 2021-11-04 | 2024-06-04 | 国网四川省电力公司营销服务中心 | Non-invasive load monitoring method and system based on die body mining and semi-supervision method |
| CN114184870A (en) * | 2021-12-14 | 2022-03-15 | 河北科技大学 | Non-invasive load identification method and equipment |
| CN114970633A (en) * | 2022-06-07 | 2022-08-30 | 西安电子科技大学 | LSTM-based non-invasive electrical appliance identification method, system and equipment |
| CN117633456A (en) * | 2023-11-17 | 2024-03-01 | 国网江苏省电力有限公司 | Marine wind power weather event identification method and device based on self-adaptive focus loss |
| CN117633456B (en) * | 2023-11-17 | 2024-05-31 | 国网江苏省电力有限公司 | Marine wind power weather event identification method and device based on self-adaptive focus loss |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112580471A (en) | Non-invasive load identification method based on AdaBoost feature extraction and RNN model | |
| CN107330517B (en) | Non-invasive resident load identification method based on S _ Kohonen | |
| CN111830347B (en) | Two-stage non-invasive load monitoring method based on event | |
| CN110907762A (en) | Non-invasive load matching identification method | |
| CN111563827B (en) | Load decomposition method based on physical characteristics of electrical appliance and electricity utilization behaviors of residents | |
| Humala et al. | Universalnilm: A semi-supervised energy disaggregation framework using general appliance models | |
| Dinesh et al. | Incorporating time-of-day usage patterns into non-intrusive load monitoring | |
| Saha et al. | Comprehensive NILM framework: Device type classification and device activity status monitoring using capsule network | |
| Chen et al. | Non-intrusive load monitoring based on feature extraction of change-point and xgboost classifier | |
| Akarslan et al. | A novel approach based on a feature selection procedure for residential load identification | |
| Cavalca et al. | Deep transfer learning-based feature extraction: An approach to improve nonintrusive load monitoring | |
| Zhang et al. | An autocorrelation incremental fuzzy clustering framework based on dynamic conditional scoring model | |
| CN113762355A (en) | User abnormal electricity consumption behavior detection method based on non-invasive load decomposition | |
| Lingqing et al. | Detection method for power theft based on SOM neural network and K-means clustering algorithm | |
| Mariño et al. | NILM: Multivariate DNN performance analysis with high frequency features | |
| Zhao et al. | Anomaly detection and classification of household electricity data: a time window and multilayer hierarchical network approach | |
| Wang et al. | Power quality disturbance recognition method in park distribution network based on one-dimensional vggnet and multi-label classification | |
| Chiam et al. | Global attention-based LSTM for noisy power quality disturbance classification | |
| Naderian | A novel hybrid deep learning approach for non-intrusive load monitoring of residential appliance based on long short term memory and convolutional neural networks | |
| Li et al. | Non-intrusive load monitoring for consistent shape loads based on convolutional neural network | |
| Jie et al. | Non-Intrusive Load Decomposition Based on Graph Convolutional Network | |
| Xiong et al. | Research on Industrial Load Decomposition Based on Correlation Analysis and DNN | |
| Liu et al. | Multi-scale residual network for energy disaggregation | |
| HU et al. | Non-intrusive load monitoring method based on SAGA-FCM algorithm | |
| Yikun et al. | Classification method of voltage sag sources based on sequential trajectory feature learning algorithm |
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: 20210330 |