CN114492671B - Non-invasive household load identification method based on improved FCM clustering algorithm and MLP neural network - Google Patents
Non-invasive household load identification method based on improved FCM clustering algorithm and MLP neural network Download PDFInfo
- Publication number
- CN114492671B CN114492671B CN202210146267.6A CN202210146267A CN114492671B CN 114492671 B CN114492671 B CN 114492671B CN 202210146267 A CN202210146267 A CN 202210146267A CN 114492671 B CN114492671 B CN 114492671B
- Authority
- CN
- China
- Prior art keywords
- load
- data
- neural network
- identification
- clustering algorithm
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 108
- 238000004422 calculation algorithm Methods 0.000 title claims abstract description 70
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 60
- 230000008569 process Effects 0.000 claims abstract description 15
- 230000006870 function Effects 0.000 claims description 56
- 238000004458 analytical method Methods 0.000 claims description 24
- 238000012549 training Methods 0.000 claims description 21
- 238000004364 calculation method Methods 0.000 claims description 20
- 230000005284 excitation Effects 0.000 claims description 20
- 238000005070 sampling Methods 0.000 claims description 19
- 230000000737 periodic effect Effects 0.000 claims description 16
- 230000008859 change Effects 0.000 claims description 13
- 210000002569 neuron Anatomy 0.000 claims description 11
- 238000003062 neural network model Methods 0.000 claims description 10
- 238000012360 testing method Methods 0.000 claims description 10
- 238000002759 z-score normalization Methods 0.000 claims description 10
- 239000011159 matrix material Substances 0.000 claims description 9
- 238000010606 normalization Methods 0.000 claims description 9
- 238000011156 evaluation Methods 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 7
- 238000002790 cross-validation Methods 0.000 claims description 6
- 230000009466 transformation Effects 0.000 claims description 4
- 238000012795 verification Methods 0.000 claims description 4
- 238000004220 aggregation Methods 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 230000001052 transient effect Effects 0.000 claims description 3
- 230000007704 transition Effects 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 11
- 238000013135 deep learning Methods 0.000 abstract description 3
- 238000004088 simulation Methods 0.000 description 12
- 238000012544 monitoring process Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 4
- 230000005611 electricity Effects 0.000 description 4
- 238000007621 cluster analysis Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 230000001174 ascending effect Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000010438 heat treatment Methods 0.000 description 2
- 238000004321 preservation Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000000630 rising effect Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000013100 final test Methods 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- QSHDDOUJBYECFT-UHFFFAOYSA-N mercury Chemical compound [Hg] QSHDDOUJBYECFT-UHFFFAOYSA-N 0.000 description 1
- 229910052753 mercury Inorganic materials 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
-
- 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
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/06—Electricity, gas or water supply
Abstract
The invention discloses a non-invasive household load identification method based on an improved FCM clustering algorithm and an MLP neural network, which comprises the following steps: carrying out household load classification and extracting power variation characteristics of the load; the FCM clustering algorithm is improved by using an entropy weight method, and the improved clustering algorithm is used for carrying out primary load identification; extracting steady-state difference current harmonic characteristics of the load; constructing an MLP neural network, and carrying out secondary identification on the load by using the trained MLP neural network; and (5) integrating the primary and secondary identification results to obtain a complete load identification result. The invention covers the data characteristics of active power, reactive power, current harmonic waves and the like of the load, the selection of various characteristics improves the accuracy of non-invasive load identification, and simultaneously, the workload of load identification only based on a deep learning method is reduced through the two identification processes, so that the accurate identification of household electric equipment can be realized, and the household load with overlapped power characteristics has good identification effect.
Description
Technical Field
The invention belongs to the field of non-invasive load monitoring, relates to a non-invasive load identification method, and in particular relates to a non-invasive household load identification method based on an improved FCM clustering algorithm and an MLP neural network.
Background
The consumption of electric energy is mainly three kinds of industrial load, commercial load and residential load. In 2020, the total social electricity consumption is 75110 hundred million kilowatt-hours, which is increased by 3.1% in the same way. The urban and rural residents use 10949 hundred million kilowatt-hours of electricity, and the electricity is increased by 6.9% in the same ratio. The power consumption of residential loads is now increasing and has become an important part of the power consumption.
In order to collect information such as power consumption of household power consumers, researchers have proposed two methods for monitoring loads of household power consumers. The first method is invasive resident power load monitoring (Intrusive Residential Load Monitoring, ILM), the method is to install a smart meter with data transmission function to all loads in the home, and then collect and process the electric energy data of electric equipment such as current, voltage and power transmitted by each meter by using the background. The second method is Non-invasive resident electric load monitoring (Non-Intrusive Load Monitoring, NILM), the method is that the intelligent ammeter is installed only at the total electric energy input port of the household electric power user, the information such as total current, voltage and power is obtained by measuring the processing ammeter, the load running states of various electric appliances of the household electric power user are obtained, and then the power consumption curve condition and the electricity consumption rule of the household electric appliances are obtained. Compared with the defects of large equipment, large investment and the like of the ILM, the NILM has low cost and easy installation and use, so that the home power consumer is more inclined to adopt a non-invasive load monitoring method. Therefore, NILM has been a popular direction of home load monitoring technology research in recent years.
Although many effective methods for non-invasive load identification have been proposed in the current research, in non-invasive load identification based on a single type of load feature and a single algorithm, the non-invasive load identification cannot have good adaptability due to the problems of overlapping features, such as high complexity of the algorithm, and the like.
Therefore, a new solution is needed to solve the above problems.
Disclosure of Invention
The invention aims to: in order to overcome the defects in the prior art, a non-invasive household load identification method based on an improved FCM clustering algorithm and an MLP neural network is provided, the data characteristics of active, reactive, current harmonics and the like of the load are covered, and the accuracy of non-invasive load identification is improved through the selection of various characteristics. Meanwhile, through the two recognition processes, the workload of load recognition based on the deep learning method is reduced, the accurate recognition of the household electric equipment is realized, and the household load with overlapped power characteristics is well recognized.
The technical scheme is as follows: in order to achieve the above object, the present invention provides a non-invasive home load identification method based on an improved FCM clustering algorithm and an MLP neural network, comprising the steps of:
S1: the type of household load is determined, and the power variation characteristic of the load is extracted;
s2: defining an entropy weight method, improving an FCM clustering algorithm by using the entropy weight method, and carrying out primary identification on the load by using the improved clustering algorithm;
s3: extracting steady-state difference current harmonic characteristics of the load;
s4: defining an MLP neural network, and carrying out secondary identification on the load by using the trained MLP neural network;
s5: and integrating the primary identification and the secondary identification results to obtain a complete load identification result.
Further, the types of the home load in the step S1 include a switch-type load, a multi-state-type load, and a continuous state-change-type load. The load with only one working state is a switch type load, including an electric kettle, an electric lamp and the like; the load with multiple working modes is a multi-state load, taking an electric cooker as an example, and the load has two working states of heating and heat preservation; the multi-state load comprises an electric cooker, an air conditioner, an electromagnetic oven and the like; the load with no significant mode switching or gentle power change is a continuous state change type load including a dimming lamp or the like.
Further, the method for extracting the power variation characteristic of the load in the step S1 is as follows:
The active change amount delta P and the reactive change amount delta Q of each load event are obtained from the obtained sampling points of the switch-type load on/off event and the multi-state load on/off/state switching event. The Δp and Δq of the sample point load event occurring at time e are:
wherein ω is the step length considering transient transition to steady state, and m is the number of sampling points selected.
Further, the core idea of the entropy weight method in step S2 is to use entropy weights to weight various data. The entropy weight represents the weight ratio among various indexes and represents the whole information of the data set. The larger the entropy weight is, the smaller the entropy value is, and the more orderly the data index is, which means that the larger the influence (weight) of the index on the system is; conversely, the smaller the entropy weight, the larger the entropy value, and the data index is disordered, meaning that the smaller the impact (weight) of this index on the system.
Further, the method for calculating the entropy weight in the step S2 is as follows:
a1: define the overall data of a system, assume X ij Is a data matrix representing the values of the j index (i=1, 2, …, n; j=1, 2, …, m) of the i-th data set.
A2: normalization of the index: the indices of a data set are typically preceded by different dimension units, which results in a failure to achieve a uniform scale between the constructed evaluation indices. Therefore, the indexes of the data set need to be normalized in advance so as to eliminate the dimensional relation among the indexes, and different index data have comparability.
There are two general normalization methods: min-max normalization and Z-score normalization. The min-max normalization method is used for carrying out linear processing on data, if outliers exist in the data, the difference of different types of data can be weakened, and entropy weight calculation is not facilitated. The Z-score normalization method can solve the problem of outliers and can be better applied to occasions with large difference of different index data [17] . Therefore, compared with the traditional entropy weight method, the method disclosed by the invention utilizes the Z-score normalization method to process data.
The original data is x 1 ,x 2 ,…,x n Changing it into standard data y with mean 0 and variance 1 i ,i＝1,2,…,n。
Wherein:is the average value of the raw data, sigma is the standard deviation of the raw data, i.e
And (3) performing Z-score normalization processing on each index data according to formulas (8) to (10) to obtain new normalized index data.
A3: calculating the proportion of the numerical value under the j index of the i data set to the index:
wherein i=1, 2, …, n; j=1, 2, …, m.
A4: calculating the entropy value of the j-th index:
wherein k=1/ln (n) > 0, e j ≥0。
A5: calculating information entropy redundancy:
d j ＝1-e j (8)
a6: calculating entropy weight of each index:
in which 0.ltoreq.ω j ≤1，
Further, the FCM clustering algorithm in step S2 is referred to as a Fuzzy C-means (FCM) clustering algorithm, and the basic principle of the algorithm is to maximize the similarity between data sets clustered to the same cluster center, while minimizing the similarity between different cluster centers.
Further, the specific principle of the FCM clustering algorithm in step S2 is as follows:
let a data set x= { F (t 1 ),F(t 2 ),…,F(t n ) Together N data sets, N being the dimension of each data set, F (t) i )∈R N Dividing X into c subsets { S ] 1 ,S 2 ,…,S c }，A＝{a 1 ,a 2 ,…,a c The set of c sub-aggregation class centers, the objective function of the FCM clustering algorithm is:
wherein U= { U ij Membership matrix of c×n, d ij For the j-th data set F (t j ) And the ith cluster center a i Of (d), i.e. d ij ＝||F(t j )-a i || 2 M is a fuzzy index, u ij Representation data set F (t j ) For the subset S where it is i Is a membership of (1).
u ij There are two constraints:
and under normal conditions, m is not lower than 1, the value of m determines the fuzzy degree of the membership matrix U, and the greater the value of m is, the greater the fuzzy degree of the clustering algorithm is, and the value range of m is (1.5, 2.5).
In order to obtain the optimal initial cluster center number, the inter-cluster entropy needs to be calculated for the result of the FCM clustering algorithm, and the inter-cluster entropy calculation formula of c clusters is as follows:
wherein,
wherein F is i 、F j Sigma for the ith and jth data set x For correlation matrices between different data sets, N is the dimension of the data set, N k The number of elements contained for each cluster. The cluster center number with the largest entropy between clusters is the optimal initial cluster center number.
Further, the method for improving the FCM clustering algorithm by using the entropy weight method in the step S2 is as follows:
the FCM algorithm performs cluster analysis by measuring the degree of difference between data objects using euclidean distance. However, the common euclidean distance is identical to the data under each index, and cannot reflect the influence degree (weight) of different indexes on the whole data. Therefore, the invention adopts the entropy weight method to carry out weighting treatment on the Euclidean distance, thereby improving the result accuracy of the FCM clustering algorithm. The euclidean distance expression adding the entropy weight is as follows.
Wherein omega is k (k=1, 2, …, L) represents the entropy weight of each index.
Further, the specific method for extracting the steady-state differential current of the load in the step S3 is as follows:
in the analysis of experimental data, it is found that, for the same type of load event, when the steady-state periodic currents (sampled from the voltage zero-crossing point of the rising trend) before and after the occurrence of the load event are respectively extracted and differenced, the waveforms of the steady-state differential currents of the load event are unchanged, i.e. the periodic current differential waveforms of the same type of load event are the same, so that the current harmonic data are also the same. The method for extracting the steady-state differential current harmonic quantity is given below.
When the load enters a steady state, the load current effective value is substantially unchanged. After the load is determined to enter a steady state, a voltage zero crossing point (voltage zero crossing point presenting an ascending situation) after the load enters the steady state is obtained through the formula (16).
U(x)＞0&U(x-1)＜0 (16)
Wherein: u (x) and U (x-1) are the sampled values of the voltage U at sampling times x and x-1, respectively.
The steady-state voltage data of one period can be obtained by the formula (16), and then the current data of the corresponding moment is determined by the period voltage data, so that the steady-state current data of one period can be obtained.
Extracting steady-state periodic currents before and after the occurrence time of the kth load event as I (t) and I (t), respectivelyObtaining a steady state differential current for the kth load event by taking the difference of equation (17):
and calculating each harmonic data of the current through Fourier transformation on the obtained steady-state difference current data. The equation for calculating steady state differential current harmonic data by discrete fourier transform (Discrete Fourier Transform, DFT) is as follows:
wherein i (N) is a steady-state differential current data sampling point of the load, and N is shared by each cycle f Sampling points from 0 to N f -1; f is the harmonic order, f=0, 1, …, N f -1; x (f) is a harmonic coefficient of DFT.
Before X (f)Multiplying by a normalized coefficient 1/N f To satisfy the condition of fourier series analysis, the euler formula is used to change the fourier series analysis into a real number calculation formula:
Wherein a is f Is the real part of X' (f), b f The imaginary part of X' (f), both of which are parameters of the f-th harmonic.
In fourier transform, the higher the number of current harmonics, the smaller the amplitude. To reduce the effects of ubiquitous noise, the current harmonic content with lower harmonic order and larger amplitude should be selected as the classification characteristic of the load event. Further, when the waveform of the periodic current satisfies the formula (22), that is, the positive half-cycle current and the negative half-cycle current of the periodic current are symmetrical about the I-t coordinate axis horizontal axis y=0, the even harmonic component and the direct current component of the periodic current are 0.
The steady state differential current waveform of the household load substantially meets the requirement of equation (22).
Therefore, the invention selects the current fundamental wave and the current harmonic amplitudes of 3, 5, 7, 9, 11 and 13 times as the characteristic of NILM secondary identification.
Further, the MLP neural network in step S4 is a common neural network model, which is composed of a plurality of superimposed sensors, and has been demonstrated to fit an arbitrary continuous function. The network structure of the MLP neural network includes 1 input layer, a plurality of hidden layers, and 1 output layer.
Further, the specific process of load identification by the MLP neural network in step S4 is as follows:
When the MLP neural network is used for load identification, the number of input nodes of the network is required to be ensured to be equal to the characteristic dimension of the load, the number of output nodes of the network is required to be equal to the classified number of categories, and each group of current harmonic data is marked with the mark of the category to which the current harmonic data belongs. The number of the input nodes of the MLP neural network constructed in the method is 7, the number of the output nodes is 1, the network structure is a single hidden layer structure, and the MLP neural network contains 4-20 neurons. The excitation function of the MLP neural network is a Sigmoid function, the cost function is a regularized cross entropy cost function, and the training algorithm is a back propagation method [19] 。
For load identification based on MLP neural network, if there are N load samples in the training set phi, the load feature dimension of each sample is M, the load class is K, then the ith load sample x i ＝[x i,1 ,x i,2 ,…,x i,M ]Load class y thereof i ＝[y i,1 ,y i,2 ,…,y i,K ]When x is i When belonging to the j-th class of load, y i,j =1, otherwise y i,j =0. Training set Φ= { (x) of network constructed according to the above method i ,y i )}。
The excitation functions f of the hidden layer and the output layer of the MLP neural network are Sigmoid functions, and the function formulas are [20] ：
z＝a T ω+b (24)
Where z is the weighted sum of the output vector a of the upper layer neuron at that neuron; omega is a weighted value; b is the weight of the +1 neuron.
In the process of constructing the MLP neural network, forward propagation calculation is firstly carried out, and load harmonic data sample x of an input layer is firstly obtained i Starting, weighting and summing, calculating excitation function output of hidden layer by layer, and taking the excitation function output as the next layerFinally, calculating the input value of the output layer excitation function result y i '。
Then, a regularized cost function is calculated, wherein the calculation formula is as follows:
in the process of calculating the cost function, the gradient of each weight without regularization term is calculated first. Based on the cost function and the excitation function, the chain rule is used for deriving, and the gradient of the output layer vector z relative to the cost function can be calculatedThen, based on a chain rule, the gradient of the last output vector a and the two-layer network weight vector omega can be calculated, and the formula is as follows:
continuously calculating the previous layer by using the chain ruleUntil the solution ends at the input layer. Then adding regularization term to the obtained network weight gradient to obtain a complete gradient about the network weight>Gradient descent methods are applied herein to calculate solutions.
The iterative formula based on gradient descent is:
wherein, alpha is a preset iterative learning rate.
After a plurality of iterative calculations, the value of the cost function is not reduced any more, and then the model is judged to be converged, and the MLP neural network model is built.
In the secondary identification of the load, steady-state difference current harmonic components of the load event to be identified are input into a network model, forward calculation is carried out, and network output is obtainedThe class in which the item with the largest value is the final identification type of the load. Here, the steady-state difference current harmonic component characteristics of the load event are used for corresponding to a class of load event, the neural network is trained through the one-to-one correspondence relationship, a group of steady-state difference current harmonic components are finally input into the network, and then the class of the load event can be obtained.
The training and testing of the final data uses a cross-validation method. Cross-validation is a very good method for verifying the effectiveness of a neural network model, and can eliminate the influence of randomly dividing training sets and test sets on model results. The method divides the original data set into K subsets, selects K-1 subsets as training sets, and the rest subsets as verification sets. Thus, K trials can be performed and K network models obtained. Taking the average of the accuracy of the K network models on the respective test sets as the final evaluation result of the neural network model.
The invention covers the data characteristics of active power, reactive power, current harmonic waves and the like of the load, and the selection of various characteristics improves the accuracy of non-invasive load identification. Meanwhile, through the two recognition processes, the workload of load recognition based on a deep learning method is reduced, the accurate recognition of household electric equipment can be realized, and the method has a good recognition effect on household loads with overlapped power characteristics. Therefore, the non-invasive household load identification method based on the improved FCM clustering algorithm and the MLP neural network has very important practical significance.
Firstly, the active and reactive variable quantity features of a load event are primarily identified once by using an FCM clustering algorithm. Then, for the load event with the power characteristics overlapped, harmonic data of the steady-state difference current is extracted. And finally, performing secondary identification on the load event by using the MLP neural network to achieve an accurate load identification effect.
According to the method, the experimental result analysis of the public data set BLUED can be realized, the household electric equipment can be accurately identified, and the method has a good identification effect on the household load with overlapped power characteristics.
The beneficial effects are that: compared with the prior art, the method has the innovation that aiming at the situation that the classification effect of the clustering algorithm is poor, the entropy weight method is used for improving the FCM clustering algorithm, so that the accuracy of load clustering is improved; the innovation of the method is that aiming at the load with overlapped characteristics, steady-state difference current harmonic components are used as the basis of load identification, and the load is further identified in a refined manner by using twice load identification, so that the accurate load identification effect is achieved.
Drawings
FIG. 1 is a flow chart of non-intrusive home load identification in accordance with the present invention;
FIG. 2 is a diagram of a portion of household load data in a BLUED data set for simulation analysis in accordance with an embodiment of the present invention;
FIG. 3 is a graph of entropy values among clusters of different clusters simulated by an embodiment of the invention;
FIG. 4 is a graph of clustering results of a conventional FCM clustering algorithm for simulation analysis according to an embodiment of the present invention;
FIG. 5 is an enlarged view of yellow cluster data simulated analysis according to an embodiment of the present invention;
FIG. 6 is a graph of the clustering results of an improved FCM clustering algorithm for simulation analysis in accordance with an embodiment of the present invention;
FIG. 7 is an enlarged view of yellow cluster data simulated analysis according to an embodiment of the present invention;
FIG. 8 is a cluster center and range diagram of a simulation analysis according to an embodiment of the present invention;
FIG. 9 is a graph of load recognition results of a conventional FCM clustering algorithm for simulation analysis according to an embodiment of the present invention;
FIG. 10 is a graph of load recognition results of an improved FCM clustering algorithm for simulation analysis in accordance with an embodiment of the present invention;
FIG. 11 is a graph of accuracy of identifying a test set by an MLP neural network for simulation analysis in accordance with an embodiment of the present invention;
FIG. 12 is a graph of load recognition results of two recognitions of a simulation analysis according to an embodiment of the present invention;
FIG. 13 is an algorithm evaluation index chart of two identifications of a simulation analysis in accordance with an embodiment of the present invention.
Detailed Description
The present invention is further illustrated in the accompanying drawings and detailed description which are to be understood as being merely illustrative of the invention and not limiting of its scope, and various modifications of the invention, which are equivalent to those skilled in the art upon reading the invention, will fall within the scope of the invention as defined in the appended claims.
The invention provides a non-invasive household load identification method based on an improved FCM clustering algorithm and an MLP neural network, referring to FIG. 1, comprising the following steps:
s1: the type of household load is determined, and the power variation characteristic of the load is extracted;
s2: defining an entropy weight method, improving an FCM clustering algorithm by using the entropy weight method, and carrying out primary load identification by using the improved clustering algorithm;
s3: extracting steady-state difference current harmonic characteristics of the load;
s4: defining an MLP neural network, and carrying out secondary identification on the load by using the trained MLP neural network;
s5: and integrating the primary identification and the secondary identification results to obtain a complete load identification result.
The types of the home load in step S1 of the present embodiment include a switch-type load, a multi-state-type load, and a continuous-state-change-type load. The load with only one working state is a switch type load, including an electric kettle, an electric lamp and the like; the load with multiple working modes is a multi-state load, taking an electric cooker as an example, and the load has two working states of heating and heat preservation; the multi-state load comprises an electric cooker, an air conditioner, an electromagnetic oven and the like; the load with no significant mode switching or gentle power change is a continuous state change type load including a dimming lamp or the like.
The method for extracting the power variation characteristic of the load in the step S1 of the embodiment is as follows:
the active change amount delta P and the reactive change amount delta Q of each load event are obtained from the obtained sampling points of the switch-type load on/off event and the multi-state load on/off/state switching event. The Δp and Δq of the sample point load event occurring at time e are:
wherein ω is the step length considering transient transition to steady state, and m is the number of sampling points selected.
The core idea of the entropy weight method in step S2 of this embodiment is to use entropy weights to perform weighting processing on various types of data. The entropy weight represents the weight ratio among various indexes and represents the whole information of the data set. The larger the entropy weight is, the smaller the entropy value is, and the more orderly the data index is, which means that the larger the influence (weight) of the index on the system is; conversely, the smaller the entropy weight, the larger the entropy value, and the data index is disordered, meaning that the smaller the impact (weight) of this index on the system.
The method for calculating the entropy weight in step S2 in this embodiment is as follows:
a1: define the overall data of a system, assume X ij Is a data matrix representing the values of the j index (i=1, 2, …, n; j=1, 2, …, m) of the i-th data set.
A2: normalization of the index: the indices of a data set are typically preceded by different dimension units, which results in a failure to achieve a uniform scale between the constructed evaluation indices. Therefore, the indexes of the data set need to be normalized in advance so as to eliminate the dimensional relation among the indexes, and different index data have comparability.
There are two general normalization methods: min-max normalization and Z-score normalization. The min-max normalization method is used for carrying out linear processing on data, if outliers exist in the data, the difference of different types of data can be weakened, and entropy weight calculation is not facilitated. The Z-score normalization method can solve the problem of outliers and can be better applied to occasions with large difference of different index data [17] . Thus, the Z-score normalization method is employed herein to process data as compared to conventional entropy weighting methods.
The original data is x 1 ,x 2 ,…,x n Changing it into standard data y with mean 0 and variance 1 i ,i＝1,2,…,n。
Wherein:is the average value of the raw data, sigma is the standard deviation of the raw data, i.e
And (3) performing Z-score normalization processing on each index data according to formulas (8) to (10) to obtain new normalized index data.
A3: calculating the proportion of the numerical value under the j index of the i data set to the index:
Wherein i=1, 2, …, n; j=1, 2, …, m.
A4: calculating the entropy value of the j-th index:
wherein k=1/ln (n) > 0, e j ≥0。
A5: calculating information entropy redundancy:
d j ＝1-e j (8)
a6: calculating entropy weight of each index:
in which 0.ltoreq.ω j ≤1，
In step S2 of this embodiment, the FCM clustering algorithm is referred to as a Fuzzy C-means (FCM) clustering algorithm, and the basic principle of the algorithm is to maximize the similarity between data sets clustered to the same cluster center, while minimizing the similarity between different cluster centers.
The specific principle of the FCM clustering algorithm in step S2 of this embodiment is as follows:
let a data set x= { F (t 1 ),F(t 2 ),…,F(t n ) Together N data sets, N being the dimension of each data set, F (t) i )∈R N Dividing X into c subsets { S ] 1 ,S 2 ,…,S c }，A＝{a 1 ,a 2 ,…,a c The aggregate of the c sub-aggregation class centers is the objective function of the FCM clustering algorithm [14] ：
Wherein U= { U ij Membership matrix of c×n, d ij For the j-th data set F (t j ) And the ith cluster center a i Of (d), i.e. d ij ＝||F(t j )-a i || 2 M is a fuzzy index, u ij Representation data set F (t j ) For the subset S where it is i Is a membership of (1).
u ij There are two constraints [14] ：
And under normal conditions, m is not lower than 1, the value of m determines the fuzzy degree of the membership matrix U, and the greater the value of m is, the greater the fuzzy degree of the clustering algorithm is, and the value range of m is (1.5, 2.5).
In order to obtain the optimal initial cluster center number, the inter-cluster entropy needs to be calculated for the result of the FCM clustering algorithm, and the inter-cluster entropy calculation formula of c clusters is as follows [15] ：
Wherein,
wherein F is i 、F j Sigma for the ith and jth data set x For correlation matrices between different data sets, N is the dimension of the data set, N k The number of elements contained for each cluster. The cluster center number with the maximum entropy between clusters is the optimal initial numberInitial cluster center number.
The method for improving the FCM clustering algorithm by using the entropy weight method in the step S2 of the embodiment is as follows:
the FCM algorithm performs cluster analysis by measuring the degree of difference between data objects using euclidean distance. However, the common euclidean distance is identical to the data under each index, and cannot reflect the influence degree (weight) of different indexes on the whole data. Therefore, the embodiment adopts the entropy weight method to carry out weighting treatment on the Euclidean distance, thereby improving the accuracy of the result of the FCM clustering algorithm. The euclidean distance expression adding the entropy weight is as follows.
Wherein omega is k (k=1, 2, …, L) represents the entropy weight of each index.
The specific method for extracting the steady-state differential current of the load in the step S3 of the embodiment comprises the following steps:
in the analysis of experimental data, it is found that, for the same type of load event, when the steady-state periodic currents (sampled from the voltage zero-crossing point of the rising trend) before and after the occurrence of the load event are respectively extracted and differenced, the waveforms of the steady-state differential currents of the load event are unchanged, i.e. the periodic current differential waveforms of the same type of load event are the same, so that the current harmonic data are also the same. The method for extracting the steady-state differential current harmonic quantity is given below.
When the load enters a steady state, the load current effective value is substantially unchanged. After the load is determined to enter a steady state, a voltage zero crossing point (voltage zero crossing point presenting an ascending situation) after the load enters the steady state is obtained through the formula (16).
U(x)＞0&U(x-1)＜0 (16)
Wherein: u (x) and U (x-1) are the sampled values of the voltage U at sampling times x and x-1, respectively.
The steady-state voltage data of one period can be obtained by the formula (16), and then the current data of the corresponding moment is determined by the period voltage data, so that the steady-state current data of one period can be obtained.
Extracting steady-state periodic currents before and after the occurrence time of the kth load event as I (t) and I (t), respectivelyObtaining a steady state differential current for the kth load event by taking the difference of equation (17):
and calculating each harmonic data of the current through Fourier transformation on the obtained steady-state difference current data. The equation for calculating steady state differential current harmonic data by discrete fourier transform (Discrete Fourier Transform, DFT) is as follows:
wherein i (N) is a steady-state differential current data sampling point of the load, and N is shared by each cycle f Sampling points from 0 to N f -1; f is the harmonic order, f=0, 1, …, N f -1; x (f) is a harmonic coefficient of DFT.
Multiplying the normalized coefficient by 1/N before X (f) f To satisfy the condition of fourier series analysis, the euler formula is used to change the fourier series analysis into a real number calculation formula:
Wherein a is f Is the real part of X' (f), b f Is the imaginary part of X' (f), both are fParameters of subharmonics.
In fourier transform, the higher the number of current harmonics, the smaller the amplitude. To reduce the effects of ubiquitous noise, the current harmonic content with lower harmonic order and larger amplitude should be selected as the classification characteristic of the load event. Further, when the waveform of the periodic current satisfies the formula (22), that is, the positive half-cycle current and the negative half-cycle current of the periodic current are symmetrical about the I-t coordinate axis horizontal axis y=0, the even harmonic component and the direct current component of the periodic current are 0.
The steady state differential current waveform of the household load substantially meets the requirement of equation (22).
Therefore, the current fundamental wave and the 3, 5, 7, 9, 11 and 13-order current harmonic amplitude are selected as the characteristic of NILM secondary identification in the embodiment.
The MLP neural network in step S4 of this embodiment is a common neural network model, which is composed of a plurality of superimposed sensors, and has been demonstrated to fit an arbitrary continuous function. The network structure of the MLP neural network includes 1 input layer, a plurality of hidden layers, and 1 output layer.
The specific process of load identification by the MLP neural network in step S4 of this embodiment is as follows:
when the MLP neural network is used for load identification, the number of input nodes of the network is required to be ensured to be equal to the characteristic dimension of the load, the number of output nodes of the network is required to be equal to the classified number of categories, and each group of current harmonic data is marked with the mark of the category to which the current harmonic data belongs. The number of input nodes of the MLP neural network constructed in the embodiment is 7, the number of output nodes is 1, the network structure is a single hidden layer structure, and the MLP neural network contains 4-20 neurons. The excitation function of the MLP neural network is a Sigmoid function, the cost function is a regularized cross entropy cost function, and the training algorithm is a back propagation method [19] 。
For the load identification based on the MLP neural network, if N load samples exist in the training set phi, the load characteristic dimension of each sample is M, and the load class is thatK, then the ith payload sample x i ＝[x i,1 ,x i,2 ,…,x i,M ]Load class y thereof i ＝[y i,1 ,y i,2 ,…,y i,K ]When x is i When belonging to the j-th class of load, y i,j =1, otherwise y i,j =0. Training set Φ= { (x) of network constructed according to the above method i ,y i )}。
The excitation functions f of the hidden layer and the output layer of the MLP neural network are Sigmoid functions, and the function formulas are [20] ：
z＝a T ω+b (24)
Where z is the weighted sum of the output vector a of the upper layer neuron at that neuron; omega is a weighted value; b is the weight of the +1 neuron.
In the process of constructing the MLP neural network, forward propagation calculation is firstly carried out, and load harmonic data sample x of an input layer is firstly obtained i Starting, weighting and summing, calculating an excitation function, calculating excitation function output of an implicit layer by layer, taking the excitation function output as an input value of a next layer, and finally calculating to obtain an excitation function result y of an output layer i '。
Then, a regularized cost function is calculated, wherein the calculation formula is as follows:
in the process of calculating the cost function, the gradient of each weight without regularization term is calculated first. Based on the cost function and the excitation function, the chain rule is used for deriving, and the gradient of the output layer vector z relative to the cost function can be calculated Then based on the chain rule, the last output vector a and the last output vector a can be calculatedThe gradient of the network weight vector omega between two layers is as follows:
continuously calculating the previous layer by using the chain ruleUntil the solution ends at the input layer. Then adding regularization term to the obtained network weight gradient to obtain a complete gradient about the network weight>Gradient descent methods are applied herein to calculate solutions.
The iterative formula based on gradient descent is:
wherein, alpha is a preset iterative learning rate.
After a plurality of iterative calculations, the value of the cost function is not reduced any more, and then the model is judged to be converged, and the MLP neural network model is built.
In the secondary identification of the load, steady-state difference current harmonic components of the load event to be identified are input into a network model, forward calculation is carried out, and network output is obtainedWherein the one with the largest valueThe class in which the item is located is the final identification type of the load.
The training and testing of the final data uses a cross-validation method. Cross-validation is a very good method for verifying the effectiveness of a neural network model, and can eliminate the influence of randomly dividing training sets and test sets on model results. The method divides the original data set into K subsets, selects K-1 subsets as training sets, and the rest subsets as verification sets. Thus, K trials can be performed and K network models obtained. Taking the average of the accuracy of the K network models on the respective test sets as the final evaluation result of the neural network model.
In step S5 of this embodiment, the complete load identification result is obtained by integrating the primary identification result and the secondary identification result of the load.
Simulation analysis:
based on the above scheme, the embodiment performs simulation verification on the MATLAB platform. The experimental analysis was performed in this example using a BLUED dataset from university of Carcinyl Mercury, which records load data for 1 consecutive week of the United states family, has power data at a sampling frequency of 60Hz and voltage current data at a sampling frequency of 12kHz, and marks the occurrence time of the load event of each electrical device [22]. The simulation test software platform uses MATLAB R2020a, and the hardware platform uses Intel i7-8550U 2.0GHz,8GB RAM computer. The experimental procedure of this embodiment selects monitoring data for 5 days for the phase a circuit of the bluetooth d data set. There are 8 common household appliances in this data: refrigerator, kitchen auxiliary knife, hairdryer, backyard lamp, air compressor, upstairs bathroom lamp, bedroom lamp and bathroom lamp. The refrigerator has 3 states, the kitchen auxiliary knife has 2 states, the bath room lamp on the building has 2 states, and the other household loads have 1 switch state and total 12 load states. Fig. 2 is a detail of a household load of a bluetooth data set.
1. Primary identification result analysis of NILM
Fig. 3 is a graph of the change of the entropy values between clusters corresponding to different numbers of clusters, and it can be seen that the entropy value between clusters is maximum when the number of clusters is 10 when the active and reactive features are used for cluster analysis, i.e. the optimal number of clusters is 10. In the experimental link of this embodiment, 12 load states are taken as a total, and since the active and reactive characteristics of the 3 load states are relatively similar, it is difficult to accurately distinguish by using the entropy value between clusters. However, it can also be seen that the use of the entropy values between clusters to determine the optimal number of clusters has enabled accurate classification of loads with distinct load characteristics, and that the detailed classification of load characteristics similarity will be studied in the next section.
Fig. 4 to 7 are graphs showing the results of classification of 12 load states by the FCM clustering algorithm before and after improvement by the entropy weight method. Wherein fig. 4 and 6 are overall cluster effect diagrams before and after improvement, fig. 5 is an enlarged view of a yellow portion of fig. 4, and fig. 7 is an enlarged view of a yellow portion of fig. 6. As can be seen from fig. 4 and fig. 5, the FCM clustering effect before improvement is not good, and the clustering center cannot be located in the center of the class, wherein the partial load states are clustered into other class load states. As can be seen from fig. 6 and fig. 7, the position of the clustering center of the FCM clustering algorithm improved by the entropy weight method is optimized, and the clustering data of the overlapping portion can be effectively distinguished.
Fig. 8 shows clustering results before and after the FCM algorithm is improved. Phase a events are classified into 10 categories altogether. Wherein, the refrigerator state 1, the refrigerator state 3, the kitchen auxiliary knife state 1, the kitchen auxiliary knife state 2, the electric hair drier, the air compressor, the upstairs bathroom lamp state 1, the bedroom lamp and the toilet lamp 9 load events are respectively gathered into one type, and the refrigerator state 2, the backyard lamp and the upstairs bathroom lamp state 2 are gathered into one type according to the steady-state active and reactive characteristics, so that the condition of overlapping of the characteristics of multiple events occurs, and the load identification is researched in the next section.
Comparing the clustering results before and after weighting in fig. 9 and fig. 10, it can be seen that the conventional FCM clustering algorithm has poor clustering effect, and 3 and 7 load events in the C2 and C3 classes are clustered into other classes, respectively. At the same time, class C6 also has 1 event clustered into other classes. The FCM algorithm improved by the entropy weight method can correct the error clusters, and the accuracy of load cluster identification is ensured.
2. Secondary identification result analysis of NILM
In the part, steady-state difference current harmonic data of three power utilization events (refrigerator state 2, backyard lamp and upstairs bathroom lamp state 2) contained in C2 class in a primary identification result are taken as input, and are verified by training of an MLP neural network built in section 2.2, and an annex A is shown for steady-state difference current waveform and harmonic data graph of each household load. Fig. 11 shows the accuracy of the MLP neural network in identifying the test set.
As can be seen from the above table, the recognition accuracy of all three devices is relatively high. The average identification accuracy of the refrigerator state 2 is highest and reaches 98.66%, the average identification accuracy of the back courtyard lamp is lower, and the accuracy is 96.83%. This is mainly because the amount of training data for different classes of devices varies in size. According to the principle of the cross validation method, 12640 sets of current harmonic data are shared by the refrigerator state 2 when the neural network training is performed every time, compared with 1800 sets of current harmonic data of the back-office lamp, the difference of the training data quantity can influence the training result accuracy of the final test set.
Fig. 12 is a diagram showing a combination of a result of the primary identification using the modified FCM algorithm and a result of the secondary identification using the MLP neural network, so that the overall non-invasive home load identification result can be intuitively seen.
The data identified in the above graph is 5 days of data in the BLUED dataset, for a total of 1288 load events that occur during the period. The primary identification correctly identified a total of 1082 load events, whereas in the secondary identification, a total of 206 load events were identified for the three power utilization events (refrigerator state 2, backyard light, and upstairs bathroom light state 2) contained in the C2 class. Wherein, there are 158 refrigerator state 2 events, 18 backyard lamp events, 30 upstairs bathroom lamp state 2 events. 156 events of the refrigerator state 2, 17 events of the backyard lamp and 28 events of the upstairs bathroom lamp state 2 can be correctly identified through the trained MLP neural network, and three load events can be basically and correctly identified.
Fig. 13 shows the results of index evaluation calculations for the fourth chapter and the two methods proposed in this chapter. The recognition accuracy of primary recognition on non-blind area load events can reach 100%, the recognition accuracy of secondary recognition on C2 type loads with overlapped features is 95.29%, the recognition accuracy of the whole secondary recognition is 98.82%, the F1 value of the secondary recognition is 97.58%, and the F1 value of the whole secondary recognition is 99.39%, so that the effectiveness and the practicability of the load recognition method provided by the invention are reflected. When the sample size of the secondary identification is further expanded, the accuracy of the load identification will be higher.
Claims (9)
1. A non-invasive family load identification method based on an improved FCM clustering algorithm and an MLP neural network is characterized by comprising the following steps:
s1: the type of household load is determined, and the power variation characteristic of the load is extracted;
s2: defining an entropy weight method, improving an FCM clustering algorithm by using the entropy weight method, and carrying out primary identification on the load by using the improved clustering algorithm;
s3: extracting steady-state difference current harmonic characteristics of the load;
s4: defining an MLP neural network, and carrying out secondary identification on the load by using the trained MLP neural network;
S5: integrating the primary identification and the secondary identification results to obtain a complete load identification result;
the definition of the entropy weight method in the step S2 is that the entropy weight is used for carrying out weighting treatment on various data; the entropy weight represents the weight ratio among various indexes and represents the whole information of the data set;
the method for calculating the entropy weight is as follows:
a1: define the overall data of a system, assume X ij A data matrix representing the j index of the i data set with the value i=1, 2, …, n; j=1, 2, …, m;
a2: normalization of the index:
data were processed using the Z-score normalization method:
the original data is x 1 ,x 2 ,…,x n Changing it into standard data y with mean 0 and variance 1 i ,i＝1,2,…,n，
Wherein:is the average value of the raw data, sigma is the standard deviation of the raw data, i.e
Performing Z-score normalization processing on each index data through formulas (8) - (10) to obtain new normalized index data;
a3: calculating the proportion of the numerical value under the j index of the i data set to the index:
wherein i=1, 2, …, n; j=1, 2, …, m;
a4: calculating the entropy value of the j-th index:
wherein k=1/ln (n) > 0, e j ≥0；
A5: calculating information entropy redundancy:
d j ＝1-e j (8)
a6: calculating entropy weight of each index:
In which 0.ltoreq.ω j ≤1，
2. The method for non-invasive home load identification based on the improved FCM clustering algorithm and the MLP neural network according to claim 1, wherein the types of home loads in step S1 include a switch-type load, a multi-state type load and a continuous state change type load.
3. The non-invasive home load identification method based on the improved FCM clustering algorithm and the MLP neural network according to claim 2, wherein the method for extracting the power variation characteristic of the load in step S1 is as follows:
the active variable quantity delta P and reactive variable quantity delta Q of each load event are obtained from the obtained sampling points of the switch-type load on/off event and the multi-state load on/off/state switching event, and the delta P and delta Q of the load event at the sampling point at the moment e are as follows:
wherein ω is the step length considering transient transition to steady state, and m is the number of sampling points selected.
4. The method for non-invasive home load identification based on the improved FCM clustering algorithm and the MLP neural network according to claim 1, wherein the method for improving the FCM clustering algorithm by using the entropy weight method in step S2 is as follows:
the Euclidean distance is weighted by adopting an entropy weight method, and the Euclidean distance expression added with the entropy weight is as follows:
Wherein omega is k (k=1, 2, …, L) represents the entropy weight of each index.
5. The non-invasive home load identification method based on the improved FCM clustering algorithm and the MLP neural network according to claim 1, wherein the method for performing the primary load identification in step S2 by the improved clustering algorithm is as follows:
the principle of the improved FCM clustering algorithm is as follows:
let a data set x= { F (t 1 ),F(t 2 ),…,F(t n ) Together N data sets, N being the dimension of each data set, F (t) i )∈R N Dividing X into c subsets { S ] 1 ,S 2 ,…,S c }，A＝{a 1 ,a 2 ,…,a c The set of c sub-aggregation class centers, the objective function of the FCM clustering algorithm is:
wherein U= { U ij Membership matrix of c×n, d ij For the j-th data set F (t j ) And the ith cluster center a i Is added with entropy weight, m is fuzzy index, u ij Representation data set F (t j ) For the subset S where it is i Membership degree of (3);
u ij there are two constraints:
m determines the degree of ambiguity of the membership matrix U;
in order to obtain the optimal initial cluster center number, the inter-cluster entropy needs to be calculated for the result of the FCM clustering algorithm, and the inter-cluster entropy calculation formula of c clusters is as follows:
wherein,
wherein F is i 、F j Sigma for the ith and jth data set x For correlation matrices between different data sets, N is the dimension of the data set, N k The number of the clustering centers with the maximum entropy among clusters is the optimal initial clustering center number for the number of elements contained in each cluster;
according to the improved FCM clustering algorithm, the active variable quantity delta P and the reactive variable quantity delta Q of the load event are used as classification basis to perform primary identification of the load.
6. The non-invasive household load identification method based on the improved FCM clustering algorithm and the MLP neural network according to claim 1, wherein the method for extracting the steady-state difference current of the load in step S3 is as follows:
when the load enters a steady state, the effective value of the load current is basically unchanged, and after the load is determined to enter the steady state, the voltage zero crossing point after the load enters the steady state is obtained through a formula (21);
U(x)＞0&U(x-1)＜0 (21)
wherein: u (x) and U (x-1) are sampling values of the voltage U at sampling moments x and x-1 respectively;
the steady-state voltage data of one period can be obtained through the formula (22), and then the current data of the corresponding moment is determined according to the period voltage data, so that the steady-state current data of one period can be obtained;
extracting steady-state periodic currents before and after the occurrence time of the kth load event as I (t) and I (t), respectivelyObtaining a steady state differential current for the kth load event by taking the difference of equation (23):
The obtained steady-state difference current data are subjected to Fourier transformation to calculate the harmonic data of the current, and the equation for calculating the steady-state difference current harmonic data through discrete Fourier transformation is as follows:
wherein i (N) is a steady-state differential current data sampling point of the load, and N is shared by each cycle f Sampling points from 0 to N f -1; f is the harmonic order, f=0, 1, …, N f -1; x (f) is the harmonic coefficient of DFT;
multiplying the normalized coefficient by 1/N before X (f) f To satisfy the condition of fourier series analysis, the euler formula is used to change the fourier series analysis into a real number calculation formula:
wherein a is f Is the real part of X' (f), b f Is the imaginary part of X' (f), both of which are parameters of the f-th harmonic;
when the waveform of the periodic current satisfies the formula (28), that is, the positive half-cycle current and the negative half-cycle current of the periodic current are symmetrical about the I-t coordinate axis y=0, the even harmonic component and the direct current component of the periodic current are 0;
the steady state differential current waveform of the household load satisfies the requirement of equation (28).
7. The method for non-invasive home load identification based on the improved FCM clustering algorithm and the MLP neural network according to claim 1, wherein the network structure of the MLP neural network in step S4 comprises 1 input layer, a plurality of hidden layers and 1 output layer.
8. The non-invasive home load identification method based on the improved FCM clustering algorithm and the MLP neural network according to claim 7, wherein the specific process of performing load secondary identification on the MLP neural network in step S4 is as follows:
when the MLP neural network is used for load identification, the number of input nodes of the network is required to be ensured to be equal to the characteristic dimension of the load, the number of output nodes of the network is required to be equal to the classified number of categories, and each group of current harmonic data is marked with a mark of the category to which the current harmonic data belongs;
the excitation function of the MLP neural network is a Sigmoid function, the cost function is a regularized cross entropy cost function, and the training algorithm is a back propagation method;
for load identification based on MLP neural network, if there are N load samples in the training set phi, the load feature dimension of each sample is M, the load class is K, then the ith negativeLoad sample x i ＝[x i,1 ,x i,2 ,…,x i,M ]Load class y thereof i ＝[y i,1 ,y i,2 ,…,y i,K ]When x is i When belonging to the j-th class of load, y i,j =1, otherwise y i,j =0; training set Φ= { (x) of network constructed according to the above method i ,y i )}；
The excitation functions f of the hidden layer and the output layer of the MLP neural network are Sigmoid functions, and the function formulas are as follows:
z＝a T ω+b (30)
where z is the weighted sum of the output vector a of the upper layer neuron at that neuron; omega is a weighted value; b is the weight of the +1 neuron;
In the process of constructing the MLP neural network, forward propagation calculation is firstly carried out, and load harmonic data sample x of an input layer is firstly obtained i Starting, weighting and summing, calculating an excitation function, calculating excitation function output of an implicit layer by layer, taking the excitation function output as an input value of a next layer, and finally calculating to obtain an excitation function result y of an output layer i '；
Then, a regularized cost function is calculated, wherein the calculation formula is as follows:
in the process of calculating the cost function, firstly, calculating the gradient of each weight without regularization item; based on the cost function and the excitation function, the chain rule is used for deriving, and the gradient V z of the output layer vector z relative to the cost function can be calculated; then, based on a chain rule, the gradient of the last output vector a and the two-layer network weight vector omega can be calculated, and the formula is as follows:
continuously calculating the V z, omega and a of the previous layer by using a chain rule until the solution is finished until the input layer is finished; the regularization term is then added to the resulting network weight gradient, resulting in a complete gradient # with respect to the network weight,
the gradient descent method is applied to calculate and solve, and an iterative formula based on gradient descent is as follows:
Wherein alpha is a preset iterative learning rate;
after iterative computation for a plurality of times, the value of the cost function is not reduced any more, and then the model is judged to be converged, and the MLP neural network model is built;
in the secondary identification of the load, steady-state difference current harmonic components of the load event to be identified are input into a network model, forward calculation is carried out, and network output is obtainedThe class in which the item with the largest value is the final identification type of the load.
9. The non-invasive home load identification method based on the improved FCM clustering algorithm and the MLP neural network according to claim 8, wherein in step S4, the final data is trained and tested by using a cross-validation method, specifically: dividing the original data set into K subsets, selecting K-1 subsets as training sets and the rest subsets as verification sets, so that K tests can be performed and K network models can be obtained, and taking the average of the accuracy of the K network models on the respective test sets as the final evaluation result of the neural network model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210146267.6A CN114492671B (en) | 2022-02-17 | 2022-02-17 | Non-invasive household load identification method based on improved FCM clustering algorithm and MLP neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210146267.6A CN114492671B (en) | 2022-02-17 | 2022-02-17 | Non-invasive household load identification method based on improved FCM clustering algorithm and MLP neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114492671A CN114492671A (en) | 2022-05-13 |
| CN114492671B true CN114492671B (en) | 2024-03-01 |
Family
ID=81482994
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210146267.6A Active CN114492671B (en) | 2022-02-17 | 2022-02-17 | Non-invasive household load identification method based on improved FCM clustering algorithm and MLP neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114492671B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116723136B (en) * | 2023-08-09 | 2023-11-03 | 南京华飞数据技术有限公司 | Network data detection method applying FCM clustering algorithm |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112329990A (en) * | 2020-10-20 | 2021-02-05 | 昆明理工大学 | User power load prediction method based on LSTM-BP neural network |
| CN112381264A (en) * | 2020-10-20 | 2021-02-19 | 昆明理工大学 | Electricity load identification method based on improved graph convolution neural network |
| CN113962314A (en) * | 2021-10-27 | 2022-01-21 | 南京富尔登科技发展有限公司 | Non-invasive enterprise load decomposition method based on federal learning |
-
2022
- 2022-02-17 CN CN202210146267.6A patent/CN114492671B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112329990A (en) * | 2020-10-20 | 2021-02-05 | 昆明理工大学 | User power load prediction method based on LSTM-BP neural network |
| CN112381264A (en) * | 2020-10-20 | 2021-02-19 | 昆明理工大学 | Electricity load identification method based on improved graph convolution neural network |
| CN113962314A (en) * | 2021-10-27 | 2022-01-21 | 南京富尔登科技发展有限公司 | Non-invasive enterprise load decomposition method based on federal learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114492671A (en) | 2022-05-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108616120B (en) | Non-invasive power load decomposition method based on RBF neural network | |
| CN111830347B (en) | Two-stage non-invasive load monitoring method based on event | |
| CN109145949A (en) | Non-intrusive electrical load monitoring and decomposition method and system based on integrated study | |
| CN107330517A (en) | One kind is based on S_Kohonen non-intrusion type resident load recognition methods | |
| CN106786534A (en) | A kind of non-intrusive electrical load transient process discrimination method and system | |
| de Paiva Penha et al. | Home appliance identification for NILM systems based on deep neural networks | |
| CN112580471A (en) | Non-invasive load identification method based on AdaBoost feature extraction and RNN model | |
| CN110416995B (en) | Non-invasive load decomposition method and device | |
| de Paiva Penha et al. | Convolutional neural network applied to the identification of residential equipment in non-intrusive load monitoring systems | |
| CN110580502A (en) | Factor hidden Markov load decomposition method based on Gaussian mixture | |
| CN109165604A (en) | The recognition methods of non-intrusion type load and its test macro based on coorinated training | |
| Li et al. | A nonintrusive load identification model based on time-frequency features fusion | |
| Jia et al. | Non-intrusive load identification using reconstructed voltage–current images | |
| CN114492671B (en) | Non-invasive household load identification method based on improved FCM clustering algorithm and MLP neural network | |
| CN114970633B (en) | LSTM-based non-invasive electrical appliance identification method, system and equipment | |
| CN111563827A (en) | Load decomposition method based on electrical appliance physical characteristics and residential electricity consumption behaviors | |
| Kang et al. | Adaptive non-intrusive load monitoring based on feature fusion | |
| CN111639586A (en) | Non-invasive load identification model construction method, load identification method and system | |
| Dash et al. | A very deep one dimensional convolutional neural network (VDOCNN) for appliance power signature classification | |
| Moradzadeh et al. | Support vector machine-assisted improvement residential load disaggregation | |
| CN114942344A (en) | Non-invasive electrical appliance identification method, system, medium, equipment and terminal | |
| Akarslan et al. | A novel approach based on a feature selection procedure for residential load identification | |
| Li et al. | Stacking ensemble learning-based load identification considering feature fusion by cyber-physical approach | |
| Bartman et al. | Identification of electrical appliances using their virtual description and data selection for non-intrusive load monitoring | |
| Deng et al. | Residential appliance detection using attention-based deep convolutional neural network |
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 |