Disclosure of Invention
The invention aims to provide a non-invasive household appliance load depth re-identification method, which solves the problem of low load identification accuracy caused by a single algorithm in the prior art.
The technical scheme adopted by the invention is that the non-invasive household appliance load depth re-identification method is implemented according to the following steps:
step 1, acquiring user high-frequency data, and performing denoising processing on the acquired user high-frequency data;
step 2, carrying out event detection on the data in the step 1 through improved generalized likelihood ratio test, if an event is detected, executing the step 3, otherwise, returning to the step 1;
step 3, extracting multi-dimensional load characteristics for the detected event change points;
step 4, aiming at the multi-dimensional features extracted in the step 3, screening relevant features by using a semi-supervised algorithm combining a Relief-F and a mRMR;
step 5, taking the characteristics obtained in the step 4 as load marks, establishing a load characteristic library through a self-adaptive FCM algorithm, identifying the load working state of the household appliances in the user by using the maximum membership degree, finishing the identification if the identification results of the two FCM algorithms are consistent, and executing the step 6 if the identification results are inconsistent;
and 6, training the parameter smoothing factor of the GRNN by adopting an SA-BAS (simulated annealing-Tianniu whisker) algorithm, and performing deep re-identification on the sample to be detected by utilizing the trained GRNN to finish the final identification of the non-invasive household appliance load.
The invention is also characterized in that:
the step 1 is implemented according to the following steps:
step 1.1, acquiring high-frequency household appliance load data containing electrical parameters, specifically representing voltage, current and corresponding power;
step 1.2, denoising processing of the power signal, wherein because an isolated noise point is easily identified as an event by an event detection algorithm, a median filtering method is selected to process the original power signal, so that the noise is eliminated without changing edge information: suppose that there is a sequence x of digital signalsj(-∞<j<+ ∞) is filtered, a window with an odd number L is first defined, where L is 2N +1, N is a positive integer, and it is assumed that at a certain time i, the signal sample in the window is xi-N,…,xi,…,xi+NWherein x isiIs the sample value of the signal located in the center of the window, and after rearranging the L signals from small to large, the value is defined as the output value of the median filter.
The step 2 is implemented according to the following steps:
step 2.1, calculating the active power P of the fundamental wave according to the formula (1)
1Taking the combined active power P as a two-dimensional power time sequence
A binary hypothesis test is proposed according to the formula (2);
in the formula, V
1Is a fundamental voltage, I
1Is the current of the fundamental wave,
is the phase difference between the two. n is
cFor the time of occurrence of the change point, k is the total length of the window, n is the last sample time in the window, μ
0,∑
0Testing for hypotheses H
0Mean of Gaussian distribution, covariance matrix, μ under the conditions
a,∑
aIs H
1Multi-dimensional signal mean, multi-dimensional covariance matrix, mu, before occurrence of change point under condition
b,∑
bIs H
1A multidimensional signal mean and a multidimensional covariance matrix after the variable point occurs under the condition;
step 2.2, defining two consecutive windows W within this time sequenceaAnd WbThe sample in the two windows is Xn={xmWhere m is n-k +1, …, n, and the window lengths are both k/2, the μ sum Σ in each of the two windows is calculated from equations (3) and (4), and then the decision function g is calculated from equation (5)n;
Step 2.3, mixing gnAnd a threshold value h1And comparing and searching suspicious points of event occurrence: when the decision function value is larger than h1When it is, refuse H0The data distribution in the two windows is not consistent, and at the point-changing time ncThere is a possibility of an event occurring; when the decision function is less than h1When it is, refuse H1And the data distribution of the two windows is consistent, and no event occurs. And when the sampling points of the detected multiple events are within 3 continuously or at intervals, the sampling points are regarded as having the maximum gnA corresponding event occurs;
and 2.4, taking the event point as a base point, and carrying out F-test screening on the false detection events. First, it is assumed that the variance values of the window data before and after the base point are equal:
calculating the value of statistic F according to equation (6), giving significance level α, rejecting hypothesis H if the value of F satisfies equation (7)
0Judging that the two maternal variances have obvious difference, and judging that an event occurs at the point;
In the formula (I), the compound is shown in the specification,
respectively the sample volume and variance of the window before and after the base point.
Step 3 is specifically carried out as follows: the method for extracting the power characteristics of the variable points specifically comprises the following steps: active power, fundamental active power, reactive power, fundamental reactive power, apparent power, distortion power, power factor angle, fundamental power factor; extracting harmonic characteristics at a variable point, specifically including voltage, amplitude of each harmonic of one to nine times of voltage, content rate of each harmonic, difference of content rates of each harmonic and total harmonic distortion rate; the current waveform characteristics comprise the wave peak value, the average value and the wave crest coefficient; the method for extracting the V-I track characteristics at the variable points specifically comprises the following steps: symmetry, surrounding direction, surrounding area, number of intersection points, Y-axis intercept, Y-axis span, midline curvature, trace middle part peak value, left and right part area, middle part shape, and instantaneous admittance standard deviation.
Step 4 is specifically implemented according to the following steps:
step 4.1, labeling part of multi-dimensional characteristic samples, and recording as marked samples S1And the remainder is marked as the unmarked sample S2The label is C, and the multidimensional feature set is A ═ A1,A2,…,AN);
Step 4.2, Slave S
1In which a sample s of class C is randomly drawn
q(C
qE.g. C); from S
2D neighboring samples are selected and recorded as
At S
1From C to C
qEach class C of the other
pRespectively solving a nearest neighbor sample x of s in the epsilon C (p is not equal to q); and at S
2D neighbor samples of medium solution x, noted
Wherein the neighbor formula is shown as formula (8); updating all feature weights based on equation (9);
in the formula, M represents the iteration number, d is the number of the adjacent samples,
for the t-th neighbor sample in the q class to which the sample s belongs,
representing the t-th neighbouring sample in the P-th class, which is different from the class of the sample s, P (C)
p) Representing the probability of the p-th class of objects, A
kAs a result of the k-th feature,
representing a sample s and a sample
With respect to feature A
kThe distance of (d);
step 4.3, circularly executing the step 4.2 for M times, and obtaining the characteristic weight omega finally output by the semi-supervised Relief-Fk'(k is 1,2, …, w'), removing features with weight coefficient less than theta to obtain a candidate feature subset, and then performing semi-supervised mRmR feature selection on the candidate feature subset;
step 4.4, marking the sample S according to the formula (11)1Calculating the correlation degree between each feature and the sample label, and obtaining the unmarked sample S according to the formula (12)2Calculating the redundancy of each characteristic;
in the formula (I), the compound is shown in the specification,
is shown in the labeled sample S
1In (1) calculation of
kMutual information with label C, R (A)
k,S
m-1) Representing a subset S of existing features
m-1Which contains m-1 features, not S
m-1Features A in the subset
kRedundancy with selected features;
step 4.5, establishing a characteristic candidate set H, and selecting the maximum correlation degree DmaxCorresponding features as candidate set leader H1Sequentially selecting the kth feature A according to equation (13)kPutting the obtained product into H until the designated characteristic number w is selected;
and 4.6, calculating each characteristic weight according to the formula (14).
Step 5 is specifically implemented according to the following steps:
step 5.1, setting the initial clustering number b, determining an initial clustering center based on a maximum-minimum distance algorithm, and simultaneously, enabling a counter k to return to zero;
a. calculating the mean of the data set
The sample point farthest from the mean is V
1j(j ═ 1,2, …, w); b. calculating the minimum distance D between each data point and the selected cluster center according to the formula (15)
xSelection of D
xThe maximum value point is used as a new clustering center; c. repeating the step b until b initial clustering centers are selected;
Dx=min d(xi,Z'k) k'=1,…,kselected (15)
step 5.2, constructing a loss function L of the FCM1 algorithm1As shown in formula (17), for uij,λjAnd vikCalculating and combining the partial derivatives, and performing u by using the derived formula (18) and formula (19)ijAnd vikThe iterative update of (2);
wherein n is the number of samples, b is the number of clusters, w is the number of features,
is a sample weight, u
ijThe membership degree of the jth sample belonging to the ith class, v is a clustering center matrix, T
uE (0, infinity) replaces the fuzzy index m for controlling the entropy value, and the maximum entropy is introduced as regularization;
step 5.3, calculating J
1A value of, if
Finishing clustering of the initial clustering number b, terminating iteration, otherwise, turning to the step 5.2 to continue iteration until the requirement is met;
step 5.4, calculating the clustering effectiveness index of the initial clustering number b according to the formula (20), namely the contour coefficient S
bIf it satisfies S
b-2<S
b-1And S
b-1>S
bIf yes, the self-adaptive FCM1 classification is ended, otherwise, whether the clustering number b reaches the maximum value or not is judged
If the maximum value is reached, all S are taken
bIf b, u and v corresponding to the medium maximum value do not reach the maximum value, b is equal to b +1, and the step 5.1 is executed again;
wherein n is the number of samples, a (i) is the average distance between the sample point i and the remaining sample objects in the same cluster, b (i) is the minimum value of the average distance between the sample point i and the sample objects in each of the remaining clusters, SbIn [ -1,1 [)]The larger the S value is, the better the clustering effect is;
step 5.5, setting the optimal clustering number b of the FCM1 as the clustering number of the FCM2, similarly adopting a maximum-minimum distance algorithm to determine an initial clustering center, and simultaneously, enabling a counter k to return to zero;
step 5.6, constructing a loss function J of the FCM2 algorithm2As shown in formula (21), for vikCalculating a partial derivative, and performing v using the derived formula (22)ikIterative updating of (1), sampling projection gradient descent method pair uijCarrying out iteration:
in the formula, lambda is more than or equal to 0 and is an orthogonal constraint parameter, the second term of the function is to add approximate orthogonal constraint to the membership matrix to balance the influence of unbalanced data, wherein u
ipThe membership degree of the ith sample point belonging to the class cluster p is represented by an original membership function u
ip' pass through
Transforming to obtain;
step 5.7, calculate J
2A value of, if
FCM2 clustering is complete and iteration terminates, otherwise k is k +1, go to step 5.6 and continue iteration until it is satisfied
Load(s)Determining the clustering number and the clustering center of the FCM algorithm in the feature library;
and 5.8, inputting a sample to be detected, determining an initial clustering center by the load feature library, obtaining a membership function through FCM clustering, taking and outputting membership vectors corresponding to the sample to be identified, sequencing the membership, taking a category corresponding to the maximum membership, and recording as the belonged load category of the sample to be identified. If the sample identification results of the FCM1 and the FCM2 are consistent, the algorithm identification is finished, the category corresponding to the maximum membership degree is the final sample identification result, and if the sample identification results are not consistent, the step 6 is executed.
Step 6 is implemented according to the following steps:
6.1, combining samples with consistent FCM identification results to serve as a training set to train the improved GRNN neural network;
step 6.2, setting parameters: the number of neurons in the input layer is a characteristic dimension w; the number n of samples in a sample set of which the pattern layer neurons are a training set, and the transfer function of the ith neuron is shown as a formula (23); the summation layer comprises two types of neurons, one is to perform arithmetic summation on the outputs of all the mode layer neurons, and the transfer function of the sum is shown as a formula (24), and the other is to perform weighted summation on the outputs of all the mode layer neurons, and the transfer function of the sum is shown as a formula (25); the number of neurons in the output layer is equal to the dimension of an output vector in a training sample, namely the number b of clustered electric appliance categories, and the network output of the electric appliance is the division of two neurons, as shown in a formula (26);
wherein X is an input sample, XiFor the ith sample data of the training set, σ is the smoothing factor, YijFor the ith output sample YiThe jth element in (a);
6.3, the network parameters of GRNN only have smooth factors, so that an SA-BAS algorithm is used for searching the optimal smooth factor sigma; firstly, inputting training data, and initializing parameters: longicorn initial position X 01, the temperature T is 200, the step factor alpha is 0.95, the maximum iteration number N is 200, the annealing cycle number L is 100, the distance d from the center of mass to the whisker is 1.5, and the counter h is 1;
step 6.4, setting the step length S of the longicorn to be T, and meanwhile, enabling a counter T to be 1;
6.5, updating the left and right beard positions of the longicorn according to a formula (27), and updating the position X of the longicorn according to a formula (28)t+1Calculating the probability according to Metropolis criterion, and judging whether to accept Xt+1As a new solution, the step S is updated as shown in equation (29)t=αSt-1Updating the distance from the center of mass to the tentacle according to the formula (30), and judging whether the counter t is more than or equal to the annealing cycle number L, if so, executing the step 6.6, otherwise, if t is t +1, and executing the step 6.5 in a circulating manner;
in the formula (I), the compound is shown in the specification,
is a random unit vector, the direction of the right hair pointing to the left hair is taken as the orientation of the longicorn in the space, sign () is a sign function, the value is greater than zero, 1 is taken, is less than zero, 1 is taken, is equal to zero, 0 is taken, T is the current temperature, and delta T is f (X)
t+1)-f(X
t),
As an expression for the fitness function, ω
jAs a feature weight, x
ijIn order to train the sample network output values,
expected output values for the training samples;
and 6.6, updating the adaptive factor beta according to the formula (31), wherein T is beta T, judging whether the counter h is more than or equal to the iteration number N, if so, outputting the optimal longicorn position as the optimal smooth factor, and finishing the training, otherwise, h is h +1, and executing the step 6.4.
In the formula (f)hFor the current fitness value, fminThe historical optimal fitness value is obtained;
and (3) building a GRNN network by combining the trained smooth factors with the conditions of the step 6.2, so that deep re-identification of the sample to be detected is performed, and final identification of the non-invasive household appliance load is completed.
The invention has the beneficial effects that: the invention discloses a non-invasive household appliance load depth re-identification method, which solves the problem of low load identification accuracy caused by a single algorithm in the prior art. The event detection algorithm combining GLR and F detection is used for locating the occurrence time of the event, the algorithm combining semi-supervised Relief-F and mRMR is used for screening out the features with high correlation with the electric label, the combination of self-adaptive FCM1 and FCM2 is used for carrying out primary judgment, the improved GRNN is used for carrying out secondary judgment, the misjudgment and feature redundancy are reduced, and the method has the advantages of high identification accuracy and high identification speed.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
The invention discloses a non-invasive household appliance load depth re-identification method, which is implemented according to the following steps as shown in figure 1:
step 1, acquiring user high-frequency data, and performing denoising processing on the acquired user high-frequency data;
the step 1 is implemented according to the following steps:
step 1.1, acquiring high-frequency household appliance load data containing electrical parameters, specifically representing voltage, current and corresponding power;
step 1.2, denoising processing of the power signal, wherein an isolated noise point is easily identified as an event by an event detection algorithm, so a median filtering method is selected to process the original power signal, and the noise is eliminated without changingEdge information: suppose that there is a sequence x of digital signalsj(-∞<j<+ ∞) is filtered, a window with an odd number L is first defined, where L is 2N +1, N is a positive integer, and it is assumed that at a certain time i, the signal sample in the window is xi-N,…,xi,…,xi+NWherein x isiIs the sample value of the signal located in the center of the window, and after rearranging the L signals from small to large, the value is defined as the output value of the median filter.
Step 2, carrying out event detection on the data in the step 1 through improved generalized likelihood ratio test, if an event is detected, executing the step 3, otherwise, returning to the step 1;
as shown in fig. 2, step 2 is specifically implemented according to the following steps:
step 2.1, calculating the active power P of the fundamental wave according to the formula (1)
1Taking the combined active power P as a two-dimensional power time sequence
A binary hypothesis test is proposed according to the formula (2);
in the formula, V
1Is a fundamental voltage, I
1Is the current of the fundamental wave,
is the phase difference between the two. n is
cFor the time of occurrence of the change point, k is the total length of the window, n is the last sample time in the window, μ
0,∑
0Testing for hypotheses H
0Mean of Gaussian distribution, covariance matrix, μ under the conditions
a,∑
aIs H
1Multi-dimensional signal mean, multi-dimensional covariance matrix, mu, before occurrence of change point under condition
b,∑
bIs H
1A multidimensional signal mean and a multidimensional covariance matrix after the variable point occurs under the condition;
step 2.2, defining two consecutive windows W within this time sequenceaAnd WbThe sample in the two windows is Xn={xmWhere m is n-k +1, …, n, and the window lengths are both k/2, the μ sum Σ in each of the two windows is calculated from equations (3) and (4), and then the decision function g is calculated from equation (5)n;
Step 2.3, mixing gnAnd a threshold value h1And comparing and searching suspicious points of event occurrence: when the decision function value is larger than h1When it is, refuse H0The data distribution in the two windows is not consistent, and at the point-changing time ncThere is a possibility of an event occurring; when the decision function is less than h1When it is, refuse H1And the data distribution of the two windows is consistent, and no event occurs. And when the sampling points of the detected multiple events are within 3 continuously or at intervals, the sampling points are regarded as having the maximum gnA corresponding event occurs;
and 2.4, taking the event point as a base point, and carrying out F-test screening on the false detection events. First, it is assumed that the variance values of the window data before and after the base point are equal:
calculating the value of statistic F according to equation (6), giving significance level α, rejecting hypothesis H if the value of F satisfies equation (7)
0Judging that the two maternal variances have obvious difference, and judging that an event occurs at the point;
In the formula (I), the compound is shown in the specification,
respectively the sample volume and variance of the window before and after the base point.
Step 3, extracting multi-dimensional load characteristics for the detected event change points;
step 3 is specifically carried out as follows: the method for extracting the power characteristics of the variable points specifically comprises the following steps: active power, fundamental active power, reactive power, fundamental reactive power, apparent power, distortion power, power factor angle, fundamental power factor; extracting harmonic characteristics at a variable point, specifically including voltage, amplitude of each harmonic of one to nine times of voltage, content rate of each harmonic, difference of content rates of each harmonic and total harmonic distortion rate; the current waveform characteristics comprise the wave peak value, the average value and the wave crest coefficient; the method for extracting the V-I track characteristics at the variable points specifically comprises the following steps: symmetry, surrounding direction, surrounding area, number of intersection points, Y-axis intercept, Y-axis span, midline curvature, trace middle part peak value, left and right part area, middle part shape, and instantaneous admittance standard deviation. Step 4, aiming at the multi-dimensional features extracted in the step 3, screening relevant features by using a semi-supervised algorithm combining a Relief-F and a mRMR;
as shown in fig. 3, step 4 is specifically implemented according to the following steps:
step 4.1, labeling part of multi-dimensional characteristic samples, and recording as marked samples S1And the remainder is marked as the unmarked sample S2The label is C, and the multidimensional feature set is A ═ A1,A2,…,AN);
Step 4.2, Slave S
1In which a sample s of class C is randomly drawn
q(C
qE.g. C); from S
2D neighboring samples are selected and recorded as
At S
1From C to C
qEach class C of the other
pRespectively solving a nearest neighbor sample x of s in the epsilon C (p is not equal to q); and at S
2D neighbor samples of medium solution x, noted
Wherein the neighbor formula is shown as formula (8); updating all feature weights based on equation (9);
in the formula, M represents the iteration number, d is the number of the adjacent samples,
for the t-th neighbor sample in the q class to which the sample s belongs,
representing the t-th neighbouring sample in the P-th class, which is different from the class of the sample s, P (C)
p) Representing the probability of the p-th class of objects, A
kAs a result of the k-th feature,
representing a sample s and a sample
With respect to feature A
kThe distance of (d);
step 4.3, circularly executing the step 4.2 for M times, and obtaining the characteristic weight omega finally output by the semi-supervised Relief-Fk'(k is 1,2, …, w'), removing features with weight coefficient less than theta to obtain a candidate feature subset, and then performing semi-supervised mRmR feature selection on the candidate feature subset;
step 4.4, marking the sample S according to the formula (11)1Calculating the correlation degree between each feature and the sample label, and obtaining the unmarked sample S according to the formula (12)2Calculating the redundancy of each characteristic;
in the formula (I), the compound is shown in the specification,
is shown in the labeled sample S
1In (1) calculation of
kMutual information with label C, R (A)
k,S
m-1) Representing a subset S of existing features
m-1Which contains m-1 features, not S
m-1Features A in the subset
kRedundancy with selected features;
step 4.5, establishing a characteristic candidate set H, and selecting the maximum correlation degree DmaxCorresponding features as candidate set leader H1Sequentially selecting the kth feature A according to equation (13)kPutting the obtained product into H until the designated characteristic number w is selected;
and 4.6, calculating each characteristic weight according to the formula (14).
Step 5, taking the characteristics obtained in the step 4 as load marks, establishing a load characteristic library through a self-adaptive FCM algorithm, identifying the load working state of the household appliances in the user by using the maximum membership degree, finishing the identification if the identification results of the two FCM algorithms are consistent, and executing the step 6 if the identification results are inconsistent;
as shown in fig. 4 and 5, step 5 is specifically implemented according to the following steps:
step 5.1, setting the initial clustering number b, determining an initial clustering center based on a maximum-minimum distance algorithm, and simultaneously, enabling a counter k to return to zero;
a. calculating the mean of the data set
The sample point farthest from the mean is V
1j(j ═ 1,2, …, w); b. calculating the minimum distance D between each data point and the selected cluster center according to the formula (15)
xSelection of D
xThe maximum value point is used as a new clustering center; c. repeating the step b until b initial clustering centers are selected;
Dx=min d(xi,Z'k) k'=1,…,kselected (15)
step 5.2, constructing a loss function L of the FCM1 algorithm1As shown in formula (17), for uij,λjAnd vikCalculating and combining the partial derivatives, and performing u by using the derived formula (18) and formula (19)ijAnd vikThe iterative update of (2);
wherein n is the number of samples, b is the number of clusters, w is the number of features,
is a sample weight, u
ijThe membership degree of the jth sample belonging to the ith class, v is a clustering center matrix, T
uE (0, infinity) replaces the fuzzy index m for controlling the entropy value, and the maximum entropy is introduced as regularization;
step 5.3, calculating J
1A value of, if
Finishing clustering of the initial clustering number b, terminating iteration, otherwise, turning to the step 5.2 to continue iteration until the requirement is met;
step 5.4, calculating the clustering effectiveness index of the initial clustering number b according to the formula (20), namely the contour coefficient S
bIf it satisfies S
b-2<S
b-1And S
b-1>S
bIf yes, the self-adaptive FCM1 classification is ended, otherwise, whether the clustering number b reaches the maximum value or not is judged
If the maximum value is reached, all S are taken
bIf b, u and v corresponding to the medium maximum value do not reach the maximum value, b is equal to b +1, and the step 5.1 is executed again;
wherein n is the number of samples, a (i) is the average distance between the sample point i and the remaining sample objects in the same cluster, b (i) is the minimum value of the average distance between the sample point i and the sample objects in each of the remaining clusters, SbIn [ -1,1 [)]In between, the larger the S value, the more polymerizedThe better the class effect;
step 5.5, setting the optimal clustering number b of the FCM1 as the clustering number of the FCM2, similarly adopting a maximum-minimum distance algorithm to determine an initial clustering center, and simultaneously, enabling a counter k to return to zero;
step 5.6, constructing a loss function J of the FCM2 algorithm2As shown in formula (21), for vikCalculating a partial derivative, and performing v using the derived formula (22)ikIterative updating of (1), sampling projection gradient descent method pair uijCarrying out iteration:
in the formula, lambda is more than or equal to 0 and is an orthogonal constraint parameter, the second term of the function is to add approximate orthogonal constraint to the membership matrix to balance the influence of unbalanced data, wherein u
ipThe membership degree of the ith sample point belonging to the class cluster p is represented by an original membership function u
ip' pass through
Transforming to obtain;
step 5.7, calculate J
2A value of, if
FCM2 clustering is complete and iteration terminates, otherwise k is k +1, go to step 5.6 and continue iteration until it is satisfied
Determining the clustering number and the clustering center of the FCM algorithm in the load characteristic library;
and 5.8, inputting a sample to be detected, determining an initial clustering center by the load feature library, obtaining a membership function through FCM clustering, taking and outputting membership vectors corresponding to the sample to be identified, sequencing the membership, taking a category corresponding to the maximum membership, and recording as the belonged load category of the sample to be identified. If the sample identification results of the FCM1 and the FCM2 are consistent, the algorithm identification is finished, the category corresponding to the maximum membership degree is the final sample identification result, and if the sample identification results are not consistent, the step 6 is executed.
And 6, training the parameter smoothing factor of the GRNN by adopting an SA-BAS algorithm, and performing secondary identification on the sample to be detected by using the trained GRNN to finish secondary identification of the non-invasive household appliance load.
As shown in fig. 6, step 6 is specifically implemented according to the following steps:
6.1, combining samples with consistent FCM identification results to serve as a training set to train the improved GRNN neural network;
step 6.2, setting parameters: the number of neurons in the input layer is a characteristic dimension w; the number n of samples in a sample set of which the pattern layer neurons are a training set, and the transfer function of the ith neuron is shown as a formula (23); the summation layer comprises two types of neurons, one is to perform arithmetic summation on the outputs of all the mode layer neurons, and the transfer function of the sum is shown as a formula (24), and the other is to perform weighted summation on the outputs of all the mode layer neurons, and the transfer function of the sum is shown as a formula (25); the number of neurons in the output layer is equal to the dimension of an output vector in a training sample, namely the number b of clustered electric appliance categories, and the network output of the electric appliance is the division of two neurons, as shown in a formula (26);
wherein X is an input sample, XiFor the ith sample data of the training set, σ is the smoothing factor, YijFor the ith output sample YiThe jth element in (a);
6.3, the network parameters of GRNN only have smooth factors, so that an SA-BAS algorithm is used for searching the optimal smooth factor sigma; firstly, inputting training data, and initializing parameters: longicorn initial position X 01, the temperature T is 200, the step factor alpha is 0.95, the maximum iteration number N is 200, the annealing cycle number L is 100, the distance d from the center of mass to the whisker is 1.5, and the counter h is 1;
step 6.4, setting the step length S of the longicorn to be T, and meanwhile, enabling a counter T to be 1;
6.5, updating the left and right beard positions of the longicorn according to a formula (27), and updating the position X of the longicorn according to a formula (28)t+1Calculating the probability according to Metropolis criterion, and judging whether to accept Xt+1As a new solution, the step S is updated as shown in equation (29)t=αSt-1Updating the distance from the center of mass to the tentacle according to the formula (30), and judging whether the counter t is more than or equal to the annealing cycle number L, if so, executing the step 6.6, otherwise, if t is t +1, and executing the step 6.5 in a circulating manner;
in the formula (I), the compound is shown in the specification,
is a random unit vector, the direction of the right hair pointing to the left hair is taken as the orientation of the longicorn in the space, sign () is a sign function, the value is greater than zero, 1 is taken, is less than zero, 1 is taken, is equal to zero, 0 is taken, T is the current temperature, and delta T is f (X)
t+1)-f(X
t),
As an expression for the fitness function, ω
jAs a feature weight, x
ijIn order to train the sample network output values,
expected output values for the training samples;
and 6.6, updating the adaptive factor beta according to the formula (31), wherein T is beta T, judging whether the counter h is more than or equal to the iteration number N, if so, outputting the optimal longicorn position as the optimal smooth factor, and finishing the training, otherwise, h is h +1, and executing the step 6.4.
In the formula (f)hFor the current fitness value, fminThe historical optimal fitness value is obtained;
and (3) building a GRNN network by combining the trained smooth factors with the conditions of the step 6.2, so that the depth re-identification of the sample to be detected is carried out, and the non-invasive secondary identification of the load of the household appliance is completed.
The invention relates to a non-invasive household appliance load depth re-identification method, which is characterized in that the occurrence time of an event is positioned through an event detection algorithm combining GLR and F detection, a semi-supervised algorithm combining Relief-F and mRmR is used for screening out characteristics with high correlation with an electric label, primary judgment is carried out through the combination of self-adaptive FCM1 and FCM2, final judgment is carried out through improved GRNN, misjudgment and characteristic redundancy are reduced, and compared with other literature methods, the result is high in identification accuracy and high in identification speed.