CN113987910A - Method and device for identifying load of residents by coupling neural network and dynamic time planning - Google Patents

Method and device for identifying load of residents by coupling neural network and dynamic time planning Download PDF

Info

Publication number
CN113987910A
CN113987910A CN202111097344.5A CN202111097344A CN113987910A CN 113987910 A CN113987910 A CN 113987910A CN 202111097344 A CN202111097344 A CN 202111097344A CN 113987910 A CN113987910 A CN 113987910A
Authority
CN
China
Prior art keywords
probability
neural network
bilstm
dtw
electric appliance
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111097344.5A
Other languages
Chinese (zh)
Inventor
滕昌志
缪巍巍
曾锃
张瑞
张明轩
李世豪
张震
张厦千
马洲俊
夏飞
张利
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Information and Telecommunication Branch of State Grid Jiangsu Electric Power Co Ltd
Original Assignee
Information and Telecommunication Branch of State Grid Jiangsu Electric Power Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Information and Telecommunication Branch of State Grid Jiangsu Electric Power Co Ltd filed Critical Information and Telecommunication Branch of State Grid Jiangsu Electric Power Co Ltd
Priority to CN202111097344.5A priority Critical patent/CN113987910A/en
Publication of CN113987910A publication Critical patent/CN113987910A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/08—Probabilistic or stochastic CAD
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/12—Timing analysis or timing optimisation

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Supply And Distribution Of Alternating Current (AREA)

Abstract

The invention provides a resident load identification method and a resident load identification device for coupling a neural network and dynamic time planning, wherein the method comprises the following steps: (1) extracting a steady-state running state; (2) extracting a gray scale image of the V-I characteristic curve; (3) extracting a feature vector; (4) training a neural network; (5) and (5) load identification. The technical field of the method is non-invasive load monitoring, the technical difficulty that a common identification method is low in identification accuracy is solved, and the method has the advantages that the selected identification features have better identification performance and smaller data dimension, and compared with the common method, the identification result has higher identification accuracy and reliability.

Description

Method and device for identifying load of residents by coupling neural network and dynamic time planning
Technical Field
The invention relates to a resident load identification method and device for coupling a neural network and dynamic time planning, and belongs to the technical field of non-invasive load identification.
Background
Load Monitoring techniques can be divided into intrusive Load Monitoring and Non-intrusive Load Monitoring (NILM).
The non-intrusive load monitoring technology is that only a measuring device needs to be installed at an inlet of an electric power system, and the types of loads in the system and the running state of the loads are monitored by collecting electrical parameters such as voltage and current of a bus.
The existing non-invasive residential load monitoring technology can be divided into a supervised learning method and an unsupervised learning method. The supervised learning method is relatively mature in development, but has the problem of low identification accuracy rate on untrained data, and is difficult to realize in engineering.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provide a resident load identification method and device for coupling a neural network and dynamic time planning, which can effectively reduce the false detection rate of an electric appliance and improve the overall identification accuracy rate.
In order to achieve the purpose, the invention is realized by adopting the following technical scheme:
in a first aspect, the present invention provides a method for identifying load of residents by coupling a neural network with dynamic time planning, which comprises the following steps:
acquiring a current and voltage time sequence of an electric appliance to be identified;
extracting the steady-state running state of the electric appliance to be identified through a probability mass function;
extracting a current and voltage time sequence of the electric appliance to be identified in a steady state operation state, carrying out normalization processing to obtain a processed sequence, and drawing a V-I characteristic curve gray scale graph according to the processed sequence;
inputting the V-I characteristic curve gray scale map into a convolutional neural network, and extracting a feature vector;
inputting the characteristic vector into a trained long-short term memory artificial neural network for identification to obtain a first probability vector set PBiLSTM(ii) a Carrying out similarity comparison with the characteristic vectors in the load database by using a dynamic time normalization algorithm to obtain a second probability vector set PDTWCoupling a first set of probability vectors PBiLSTMWith a second set of probability vectors PDTWAnd obtaining a final identification result.
Further, the construction method of the load database comprises the following steps:
the construction method of the load database comprises the following steps:
collecting current and voltage time sequences of various electrical appliances, and extracting steady-state operation states of the various electrical appliances through a probability mass function;
extracting current and voltage time sequences of various electrical appliances in a steady-state operation state, carrying out normalization processing to obtain processed sequences of the various electrical appliances, and drawing a V-I characteristic curve gray scale diagram of the various electrical appliances according to the processed sequences of the various electrical appliances;
inputting the V-I characteristic curve gray level graphs of various electrical appliances into a convolutional neural network, and extracting characteristic vectors of the various electrical appliances;
the feature vectors of the various appliances are combined to form a load database.
Further, the method for extracting the steady-state operation state through the probability mass function PMF comprises the following steps:
the calculation is as follows:
PMis a discrete active power sequence of the electric appliance M, wherein the maximum power value is Pmax(ii) a Setting the classification power interval to 20W, and dividing PmaxGrading according to the grading power interval and counting PMThe number Num of times of power appearing in each stage is calculated, and P is calculatedMThe classification probability of the occurrence of the power of each stage;
if the classification probability of a certain level is simultaneously greater than the classification probability of the previous level and the classification probability of the next level, the classification probability of the level is considered as a PMF peak value, and the active power classification range corresponding to the PMF peak value is a steady-state operation state of the electric appliance.
Further, the method for obtaining the processed sequence by normalizing the current and voltage time sequence in the steady-state operation state comprises the following steps:
the current and voltage time sequence of the steady state waveform of each period is taken, one period is taken as a data points, the current and voltage data are converted into uniform size of 0-l through normalization, and the formula is as follows:
Figure BDA0003269563980000031
Figure BDA0003269563980000032
in the formula, bcAnd ucC sampling point data representing normalized current and voltage; i iscAnd UcThe c-th sampling point data which represents the original current and voltage; minI and maxI respectively represent the minimum value and the maximum value of the primary current in one periodA value; minU and maxU respectively represent the minimum value and the maximum value of the original voltage in one period; []The integer symbol is expressed, and l is a fixed value, namely, the current and voltage data are converted into a fixed value of l.
Further, the method for drawing the V-I characteristic curve gray scale map according to the processed sequence comprises the following steps:
respectively taking the current and voltage data of the waveform of the electrical appliance in the steady-state operation of each period in the processed sequence as a vertical coordinate and a horizontal coordinate to draw a V-I characteristic curve;
converting the V-I characteristic curve into a 47-by-47 matrix, wherein matrix elements of the matrix range from 0 to 255; and representing 256 gray levels of the image by using the matrix elements of the matrix, and converting the gray levels into a V-I characteristic curve gray level map in the form of a gray level picture.
Further, the structure of the convolutional neural network comprises:
an input layer with dimension (47, 47, 1);
a convolution layer with dimensions (44, 44, 20), a kernel function dimension of 5 × 5(20), and an activation function of tanh;
a layer of average pooling layers of dimensions (22, 22, 20), kernel function dimensions of 2 × 2;
a fully-connected layer with the dimension of (1, 100) and an activation function of Relu;
the feature vectors extracted by the convolutional neural network are feature vectors of size 1 x 100.
Further, the structure of the long-short term memory artificial neural network comprises: an input layer with dimension (1, 100), a BilSTM layer with dimension (1, 100), wherein the activation functions are sigmoid and tanh; a fully-connected layer with dimension (1, 11); a classification layer with dimension (1, 9), wherein the activation function is Softmax;
the training method of the long-short term memory artificial neural network comprises the following steps:
inputting the electric appliance characteristic vectors in the load database into a long-term and short-term memory artificial neural network for training;
in the training process, a gradient threshold value is set to be 1, a training period is set to be 60, each period is iterated for 54 times, and the training process uses an adaptive moment estimation optimization algorithm.
Further, a first set of probability vectors P is coupledBiLSTMWith a second set of probability vectors PDTWThe method for obtaining the final identification result comprises the following steps:
combining the probability vector sets of the two algorithms through the following formula to form a final identification vector set Pfinal＝{Pfinal(1),Pfinal(2),…,Pfinal(i)…,Pfinal(k)}：
Figure BDA0003269563980000041
In the formula, Pfinal(i) Representing the final probability value of the electric appliance to be identified as an electric appliance i, wherein the value range of i is 1 to k, k is the total number of the electric appliances which can be identified in the database, and P is the total number of the electric appliances which can be identified in the databaseBiLSTM＝{PBiLSTM(1),PBiLSTM(2),…,PBiLSTM(i)…,PBiLSTM(k) Represents the set of probability values, P, of the output of the BILSTM modelBiLSTM(i) Representing the probability value of the electric appliance i identified by the CNN-BilSTM network; pDTW＝{PDTW(1),PDTW(2),…,PDTW(i)…,PDTW(k) Represents the DTW probability vector set, P, of the DTW algorithm outputDTW(i) Representing the probability value of the electric appliance i identified by the DTW algorithm, wherein the value range of i is 1 to k, and k is the total number of the electric appliances which can be identified in the database;
comparing the final set of discriminatory vectors PfinalThe maximum element in the set is selected according to the size of each element in the electric appliance to be identified, and then the final identification result of the electric appliance to be identified can be obtained.
In a second aspect, the present invention provides a device for identifying load of residents coupling a neural network with dynamic time planning, the device comprising:
an acquisition module: the method comprises the steps of obtaining a current and voltage time sequence of an electric appliance to be identified;
a steady state extraction module: the device is used for extracting the steady-state operation state of the electric appliance to be identified through the probability mass function;
a curve extraction module: extracting a current and voltage time sequence of the electric appliance to be identified in a steady state operation state, carrying out normalization processing to obtain a processed sequence, and drawing a V-I characteristic curve gray scale graph according to the processed sequence;
the vector extraction module: the system is used for inputting the V-I characteristic curve gray scale map into a convolution neural network and extracting a characteristic vector;
identifying the coupling module: the characteristic vectors are input into a trained long-short term memory artificial neural network for identification to obtain a first probability vector set PBiLSTM(ii) a Carrying out similarity comparison with the characteristic vectors in the load database by using a dynamic time normalization algorithm to obtain a second probability vector set PDTWCoupling a first set of probability vectors PBiLSTMWith a second set of probability vectors PDTWAnd obtaining a final identification result.
Compared with the prior art, the invention has the following beneficial effects:
1. the invention respectively calculates the first probability vector set and the second probability vector set of the long-short term memory artificial neural network and the dynamic time reduction algorithm, and the identification mechanisms of the long-short term memory artificial neural network and the dynamic time reduction algorithm are greatly different, so that the types of electrical appliances which are well identified by the long-short term memory artificial neural network and the dynamic time reduction algorithm have larger difference, the identification results of the long-short term memory artificial neural network and the dynamic time reduction algorithm are coupled, the false detection rate of the electrical appliances can be reduced, and the integral identification accuracy rate can be improved;
2. compared with other load marks, the load marks formed by extracting the characteristic vectors of the V-I characteristic curve gray level graph by using the convolutional neural network CNN have better identification performance, and the data dimension is smaller;
3. compared with other algorithms, the deep learning-based long-short term memory artificial neural network has higher identification accuracy and higher reliability of the result.
Drawings
FIG. 1 is a flow chart of the present invention.
Fig. 2 is a gray scale diagram of V-I characteristic curve of a common household appliance.
FIG. 3 is a diagram of the structure of the LSTM.
FIG. 4 is a comparison of recognition accuracy for different algorithms.
Detailed Description
The invention is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
The first embodiment is as follows:
the present embodiment provides a method for identifying load of residents by coupling a neural network and dynamic time planning, a flowchart of which is shown in fig. 1, and the method comprises the following steps:
step 1: extracting a steady-state operation state: collecting voltage and current data of the household appliance, and extracting a steady-state operation state through a probability mass function. In this embodiment, the steady-state operation state included in the electrical appliance is counted by calculating the probability quality function of the active power of the electrical appliance, and a mathematical model thereof is as follows.
The discrete active power sequence of a certain electric appliance M is PM＝{P1,P2,...,Pd,…,PnAt a maximum power value Pmax. Setting the classification power interval to 20W (where the classification power interval may be different settings, but according to experimental experience, 20W is just about right, too small will make the disturbance effect large, and too large will make the state error of the small power electric appliance too large), according to PmaxRank and count PMThe number Num of the power occurrences of each stage can be obtained by the following formula:
Figure BDA0003269563980000061
in the formula, PdRepresents the sequence PMPro (d) represents the probability that the appliance is operating at the d power level; num (d) represents the total number of values in the historical discrete power at the d-th power level, and n is the number of elements of the discrete active power sequence.
After the grading probability is counted, if Pro (d) is greater than both Pro (d-1) and Pro (d +1), Pro (d) is considered to be a PMF peak value, and the corresponding active power grading range is a steady-state operation state of the electric appliance.
Step 2: and (3) extracting a V-I characteristic curve gray level graph: the method comprises the steps of extracting current and voltage time sequences of each electric appliance under the steady-state operation, carrying out normalization processing, drawing a bmp-format gray level picture with a V-I characteristic curve in a uniform scale, namely a gray level picture of the V-I characteristic curve.
The current and voltage time sequence of the steady state waveform of each period is taken, one period is taken as a data points, the current and voltage data are converted into uniform size of 0-l through normalization, and the formula is as follows:
Figure BDA0003269563980000071
Figure BDA0003269563980000072
in the formula, bcAnd ucC sampling point data representing normalized current and voltage; i iscAnd UcThe c-th sampling point data which represents the original current and voltage; minI and maxI respectively represent the minimum value and the maximum value of the original current in one period; minU and maxU respectively represent the minimum value and the maximum value of the original voltage in one period; []The integer symbol is expressed, and l is a fixed value, namely, the current and voltage data are converted into a fixed value of l.
The voltage is used as an abscissa and the current is used as an ordinate to draw a V-I characteristic curve, the V-I characteristic curve can be regarded as an l multiplied by l matrix, the value range of elements in the matrix is 0 to 255, and the 256 gray levels of an image are represented by the value of the elements of the matrix to form a gray level graph, namely the gray level graph of the V-I characteristic curve.
Preferably, in order to make the feature vector obtained by CNN extraction more recognizable, the gray scale of each pixel in the gray scale map is converted, if the gray scale is less than 250, the gray scale is converted into 255, and the remaining pixels are all converted into 0, and the gray scale map of the V-I characteristic curve of the common household appliance is shown in fig. 2.
And step 3: extracting the feature vector: and (3) inputting the gray scale map of the V-I characteristic curve obtained in the step (2) into a Convolutional Neural Network (CNN), extracting the characteristic vector of the electric appliance, and forming a load database.
The convolutional neural network CNN is one of deep neural network structures, is usually used for extracting picture characteristic information, and has great advantages in processing of two-dimensional data.
In general, the convolutional neural network CNN can be classified into: input layer, hidden layer, output layer. The hidden layer can be divided into a convolutional layer (convolutional layer), a pooling layer (posing layer), and a fully-connected layer (fully-connected layer). Neurons in each layer are connected to each other, and neurons in the same layer are not connected.
Wherein the convolutional layer can be defined by the following formula:
Figure BDA0003269563980000081
in the formula, Xj nRepresenting the jth characteristic picture of the nth layer; kij nA convolution kernel function between the jth characteristic diagram representing n layers and the n-1 st layer; bj nRepresents the jth offset number of the nth layer; f is an activation function for converting the convolution result into a nonlinear output.
Wherein the pooling layer may be defined by the following formula:
Figure BDA0003269563980000082
in the formula, Xj nRepresenting the jth characteristic picture of the nth layer; alpha is alphaj nRepresents the jth weight of the nth layer; beta is aj nRepresents the jth offset number of the nth layer; pool represents a pooling function, typically having a zone maximum function or a zone mean function; f is the activation function.
And after the convolution layer and the pooling layer are processed, mapping the output data to a classification mark space through the full connection layer, and finally converting the output of the full connection layer into a classification mark by using a softmax function. In the invention, the output of the full connection layer is intercepted, and a flat operation layer is added to convert the matrix into an array for the subsequent algorithm reading.
In the embodiment, the CNN is used for extracting the information of the load V-I characteristic curve, the load V-I characteristic curve is converted into a one-dimensional characteristic array from a two-dimensional gray level picture, the impedance characteristic information of the electric appliance is retained to the maximum extent, and the impedance characteristic information is converted into an information format which is easy to read by a neural network.
Specifically, the method for constructing the load database of the present embodiment includes the following steps:
collecting current and voltage time sequences of various electrical appliances, and extracting steady-state operation states of the various electrical appliances through a probability mass function;
extracting current and voltage time sequences of various electrical appliances in a steady-state operation state, carrying out normalization processing to obtain processed sequences of the various electrical appliances, and drawing a V-I characteristic curve gray scale diagram of the various electrical appliances according to the processed sequences of the various electrical appliances;
inputting the V-I characteristic curve gray level graphs of various electrical appliances into a convolutional neural network, and extracting characteristic vectors of the various electrical appliances;
the feature vectors of the various appliances are combined to form a load database.
And 4, step 4: training a neural network: and inputting a load database consisting of the characteristic vectors of the electric appliances into the long-term and short-term memory artificial neural network for training.
The long-short term memory artificial neural network is one kind of cyclic neural network, solves the problem that the gradient disappears during RNN training, and has good performance when processing long-term sequences. The LSTM is composed of a plurality of identical cell structures, each cell structure is composed of a forgetting gate, an input gate, and an output gate, and the topology is as shown in fig. 3:
in FIG. 3, CtInformation indicating the state of the cells at time t, htCell output information, x, at time ttCell input information at time t is shown. f. oftAn output representing a forgetting gate at time t, whichThe operation can be represented by the following formula:
ft＝σ(Wf·[ht-1,xt]+bf)
wherein σ is an activation function sigmoid, WfAnd bfRespectively represent the weight and bias of the forgetting gate, [ h ]t-1,xt]Representing the union of the input at time t and the output at the previous time.
itAnd
Figure BDA0003269563980000091
collectively representing the output of the input gates, the operation of which can be represented by the following equation:
it＝σ(wi·[ht-1,xt]+bi)
Figure BDA0003269563980000092
wherein σ and tanh represent activation functions, WiAnd WCRepresents a weight, biAnd bCIndicating the bias.
The operation of cell state update can be represented by the following formula:
Figure BDA0003269563980000101
the operation of the final output gate can be represented by the following equation:
Ot＝σ(WO[ht-1,xt]+bO)
ht＝Ot·tanh(Ct) (9)
in the formula, OtRepresenting output gate intermediate information, sigma and tanh representing activation functions, WORepresents a weight, bOIndicating the bias.
The BilSTM is a variant algorithm of the LSTM and comprises two layers of LSTMs, wherein one layer of LSTM is propagated forwards, the other layer of LSTM is propagated backwards, the forward layer starts input iteration from the starting point of the number sequence, the backward layer starts input iteration from the tail end of the number sequence, and finally the output results of the two layers are fitted to obtain an identification result.
BilSTM can be expressed by the following formula:
ht＝f(W1xt+W2ht-1)
rt＝f(W3xt+W5rt+1)
yt＝g(W4ht+W6rt)
in the formula, xtRepresenting the input quantity at time t; h istAnd rtRespectively representing the output quantities of a forward layer and a reverse layer at the time t; y istRepresents the output of the output layer at time t; f represents the activation function of the forward layer and the reverse layer; g represents an activation function of the output layer; w1And W3Mapping the input layer to a weight matrix of a forward layer and a reverse layer; w2And W5Outputting a weight matrix mapped to the current calculation time for the previous calculation time of the forward layer and the reverse layer; w4And W6A weight matrix that maps outputs of the forward layer and the reverse layer to an output layer.
And 5: load identification: extracting a steady state operation state of an electric appliance to be identified, then obtaining a V-I characteristic curve gray scale graph in the steady state operation state, extracting a feature vector of the V-I characteristic curve gray scale graph by using CNN, and then inputting the feature vector into a trained neural network for identification to obtain a probability vector PBiLSTM(ii) a Meanwhile, the DTW is used for carrying out similarity comparison with the feature vectors in the load database to obtain the probability vector P with the maximum similarityDTWIs coupled with PBiLSTMAnd PDTWAnd obtaining a final identification result.
The mathematical model of the DTW algorithm is summarized as follows: two time series signals A ═ a are set1,a2,…,anB ═ B } and B ═ B1,b2,…,bm}. Firstly, a cost matrix D between A and B is calculated, wherein the matrix is an n multiplied by m order matrix, and the expression of the matrix is shown as the following formula:
Figure BDA0003269563980000111
in the formula (d)nmDenotes anAnd bmOf between, wherein dnm＝||an-bm||2，||||2Representing a 2 norm.
Then, d in the cost matrix is found11To dnmTo minimize the sum of D on the path, and construct a new cost matrix DdistWherein the element distijAs shown in the following equation:
Figure BDA0003269563980000112
i.e. the weighted sum of the shortest local cost measure between a and B is the cost matrix DdistElement dist ofnmThe value of (c).
Because the difference between the identification mechanisms of the BilSTM and the DTW is huge, the BilSTM and the DTW are different from each other for electrical appliances with higher identification degrees, the final identification accuracy can be effectively improved by combining the identification results of the two algorithms, and meanwhile, the range of the identifiable electrical appliances can be expanded.
Inputting the characteristic vector into a trained long-short term memory artificial neural network for identification to obtain a BilSTM probability vector set PBiLSTM(ii) a Meanwhile, similarity comparison is carried out between the Dynamic Time Warping (DTW) algorithm and the feature vectors in the load database to obtain a DTW probability vector set PDTW。
PBiLSTM＝{PBiLSTM(1),PBiLSTM(2),…,PBiLSTM(i)…,PBiLSTM(k) Represents the set of probability values, P, of the output of the BILSTM modelBiLSTM(i) Representing the probability value of the electric appliance i identified by the CNN-BilSTM network;
PDTW＝{PDTW(1),PDTW(2),…,PDTW(i)…,PDTW(k) represents the DTW probability vector set, P, of the DTW algorithm outputDTW(i) Stand for standingAnd identifying the probability value of the electrical appliance i identified by the DTW, wherein the value range of i is 1 to k, and k is the total number of the identifiable electrical appliances in the database of the two algorithms.
Combining the probability vector sets of the two algorithms through the following formula to form a final identification vector set Pfinal＝{Pfinal(1),Pfinal(2),…,Pfinal(i)…,Pfinal(k)}：
Figure BDA0003269563980000121
In the formula, Pfinal(i) Representing the final probability value of the electric appliance to be identified as an electric appliance i, wherein the value range of i is 1 to k, k is the total number of the electric appliances which can be identified in the database, and P is the total number of the electric appliances which can be identified in the databaseBiLSTM＝{PBiLSTM(1),PBiLSTM(2),…,PBiLSTM(i)…,PBiLSTM(k) Represents the set of probability values, P, of the output of the BILSTM modelBiLSTM(i) Representing the probability value of the electric appliance i identified by the CNN-BilSTM network; pDTW＝{PDTW(1),PDTW(2),…,PDTW(i)…,PDTW(k) Represents the DTW probability vector set, P, of the DTW algorithm outputDTW(i) Representing the probability value of the electric appliance i identified by the DTW, wherein the value range of i is 1 to k, and k is the total number of the electric appliances which can be identified in the database;
comparing the final set of discriminatory vectors PfinalThe maximum element in the set is selected according to the size of each element in the electric appliance to be identified, and then the final identification result of the electric appliance to be identified can be obtained.
Namely, the largest element in the BILSTM probability vector set output by the BILSTM model and the DTW probability vector set output by the DTW algorithm is screened out, and the largest element is the final identification result of the electric appliance to be identified. It should be noted that there are many methods for selecting the largest element in the two probability vector sets, and the implementation method of the present invention is not limited to this method.
In particular, PBiLSTM(i) And PDTW(i) Respectively represents that the electric appliances to be identified are identified by two algorithmsThe probability of a certain electrical appliance i is calculated according to the following formula:
PBiLSTM(i)＝Softmax(W4(hi)+W6(ri))
Figure BDA0003269563980000122
wherein i is the identified electrical appliance; softmax is the activation function; w4And W6Respectively outputting a weight matrix which is mapped to an output layer by a BiLSTM forward layer and a reverse layer; h isi、riRespectively outputting the output quantity of the output neuron of the electrical appliance i in the database corresponding to the forward layer and the reverse layer; dist (i) represents the DTW distance calculated between the characteristics of the electrical appliance to be identified and the electrical appliance i, the smaller the distance is, the more similar the electrical appliance to be identified is, and Σ dist (k) represents the sum of the distances between the electrical appliance to be identified and each electrical appliance.
For practical situations of the invention, the following simulation experiments were designed for verification:
the characteristic vector of the V-I characteristic curve gray scale graph extracted by the CNN is used as the load mark, and in order to verify the effectiveness of the load mark and higher identifiability relative to other load marks, the load mark selected by the invention and other various load marks such as the V-I characteristic curve gray scale graph, active power and reactive power, current and voltage, current harmonic waves and the like are compared and analyzed under a BilSTM model.
Figure BDA0003269563980000131
Table 1 shows a summary table of the recognition accuracy rates when different loads are used for marking under the same BilSTM training model. Experiments show that when the gray-scale graph of the V-I characteristic curve is used as the load mark, the identification accuracy of the algorithm is higher than that of the load mark except the text, and the identification accuracy reaches 92.7% after the load mark is adopted and is higher than that of the gray-scale graph of the V-I characteristic curve directly used, so that the characteristic vector after CNN conversion not only enables the training data to be reduced in dimension, reduces the training complexity, but also improves the overall load identification accuracy.
In order to further verify the superiority of the algorithm, a plurality of common classification algorithms are compared with the algorithm of the invention in identification accuracy, a PLAID public data set is adopted as experimental data, each algorithm is tested repeatedly for 100 times, and the result is shown in figure 4; the Plug-load application Identification Dataset (PLAID) is the name of a public data set containing voltage and current measurements from different household appliances, with a sampling frequency of 30kHz, collected in 65 different locations in pittsburgh, pa. The arithmetic mean of the overall recognition accuracy is compared, and the result is shown in Table 2, wherein K-means represents K-means Clustering Algorithm (K-means Clustering Algorithm), and GMM represents Gaussian mixture Model (Gaussian mixture Model).
The analysis of the attached figure 4 shows that the algorithm identification accuracy of the neural network is high in reliability, the result fluctuation is small under 100 tests, the reliability is higher compared with other algorithms, and the method has better reproducibility in practical application. The algorithm has the highest identification accuracy and better identification performance.
Figure BDA0003269563980000141
As can be seen from the analysis table 2, the effect of the BilSTM on the identification of the electric appliance is better, and the average identification accuracy rate obtained after 100 times of tests is 91.7 percent and is higher than that of the k-means algorithm and the GMM algorithm; after the CNN and the BilSTM are coupled into a new network, the average identification accuracy reaches 92.5 percent and is higher than the average value of the accuracy of the two algorithms during independent test, and the identification accuracy can be effectively improved by using the BilSTM identification after the characteristic vector is extracted by the CNN. Finally, the average accuracy of the algorithm is 94.1%, which is higher than that of the comparison algorithm, and experiments prove that the algorithm effectively improves the accuracy of family load identification.
Example two:
the embodiment provides a resident load identification device coupling a neural network and dynamic time planning, the device comprises:
an acquisition module: the method comprises the steps of obtaining a current and voltage time sequence of an electric appliance to be identified;
a steady state extraction module: the device is used for extracting the steady-state operation state of the electric appliance to be identified through the probability mass function;
a curve extraction module: extracting a current and voltage time sequence of the electric appliance to be identified in a steady state operation state, carrying out normalization processing to obtain a processed sequence, and drawing a V-I characteristic curve gray scale graph according to the processed sequence;
the vector extraction module: the system is used for inputting the V-I characteristic curve gray scale map into a convolution neural network and extracting a characteristic vector;
identifying the coupling module: the characteristic vector is input into a trained long-short term memory artificial neural network for identification to obtain a first probability vector PBiLSTM(ii) a Meanwhile, the similarity comparison is carried out between the dynamic time normalization algorithm and the characteristic vectors in the load database to obtain a second probability vector PDTWCoupled to the first probability vector PBiLSTMAnd a second probability vector PDTWObtaining the final identification result Pfinal。
The apparatus of the present embodiment can be used to implement the method described in the first embodiment.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should also be regarded as the protection scope of the present invention.

Claims (9)

1. A resident load identification method for coupling a neural network and dynamic time planning is characterized by comprising the following steps:
acquiring a current and voltage time sequence of an electric appliance to be identified;
extracting the steady-state running state of the electric appliance to be identified through a probability mass function;
extracting a current and voltage time sequence of the electric appliance to be identified in a steady state operation state, carrying out normalization processing to obtain a processed sequence, and drawing a V-I characteristic curve gray scale graph according to the processed sequence;
inputting the V-I characteristic curve gray scale map into a convolutional neural network, and extracting a feature vector;
inputting the characteristic vector into a trained long-short term memory artificial neural network for identification to obtain a first probability vector set PBiLSTM(ii) a Carrying out similarity comparison with the characteristic vectors in the load database by using a dynamic time normalization algorithm to obtain a second probability vector set PDTWCoupling a first set of probability vectors PBiLSTMWith a second set of probability vectors PDTWAnd obtaining a final identification result.
2. The method for identifying loads of residents coupling neural network and dynamic time planning as claimed in claim 1, wherein the method for constructing the load database comprises the following steps:
collecting current and voltage time sequences of various electrical appliances, and extracting steady-state operation states of the various electrical appliances through a probability mass function;
extracting current and voltage time sequences of various electrical appliances in a steady-state operation state, carrying out normalization processing to obtain processed sequences of the various electrical appliances, and drawing a V-I characteristic curve gray scale diagram of the various electrical appliances according to the processed sequences of the various electrical appliances;
inputting the V-I characteristic curve gray level graphs of various electrical appliances into a convolutional neural network, and extracting characteristic vectors of the various electrical appliances;
the feature vectors of the various appliances are combined to form a load database.
3. The method for identifying loads of residents coupling a neural network with dynamic time planning as recited in claim 2, wherein the method for extracting the steady-state operation state through the probability mass function PMF comprises:
the calculation is as follows:
PMis a discrete active power sequence of the electric appliance M, wherein the maximum power value is Pmax(ii) a Setting the classification power interval to 20W, and dividing PmaxGrading according to the grading power interval and counting PMThe number Num of times of power appearing in each stage is calculated, and P is calculatedMThe classification probability of the occurrence of the power of each stage;
if the classification probability of a certain level is simultaneously greater than the classification probability of the previous level and the classification probability of the next level, the classification probability of the level is considered as a PMF peak value, and the active power classification range corresponding to the PMF peak value is a steady-state operation state of the electric appliance.
4. The method for identifying loads of residents coupling a neural network and dynamic time planning as claimed in claim 2, wherein the current and voltage time series under the steady state operation state are normalized, and the method for obtaining the processed series comprises:
the current and voltage time sequence of the steady state waveform of each period is taken, one period is taken as a data points, the current and voltage data are converted into uniform size of 0-l through normalization, and the formula is as follows:
Figure FDA0003269563970000021
Figure FDA0003269563970000022
in the formula, bcAnd ucC sampling point data representing normalized current and voltage; i iscAnd UcThe c-th sampling point data which represents the original current and voltage; minI and maxI respectively represent the minimum value and the maximum value of the original current in one period; minU and maxU respectively represent the minimum value and the maximum value of the original voltage in one period; []The integer symbol is expressed, and l is a fixed value, namely, the current and voltage data are converted into a fixed value of l.
5. The method for identifying loads of residents coupling a neural network with dynamic time planning as claimed in claim 2, wherein the method for drawing the gray scale map of the V-I characteristic curve according to the processed sequence comprises the following steps:
respectively taking the current and voltage data of the waveform of the electrical appliance in the steady-state operation of each period in the processed sequence as a vertical coordinate and a horizontal coordinate to draw a V-I characteristic curve;
converting the V-I characteristic curve into a 47-by-47 matrix, wherein matrix elements of the matrix range from 0 to 255; and representing 256 gray levels of the image by using the matrix elements of the matrix, and converting the gray levels into a V-I characteristic curve gray level map in the form of a gray level picture.
6. The method for identifying load of residents coupling a neural network with dynamic time planning as claimed in claim 2, wherein said convolutional neural network is structured as follows:
an input layer with dimension (47, 47, 1);
a convolution layer with dimensions (44, 44, 20), a kernel function dimension of 5 × 5(20), and an activation function of tanh;
a layer of average pooling layers of dimensions (22, 22, 20), kernel function dimensions of 2 × 2;
a fully-connected layer with the dimension of (1, 100) and an activation function of Relu;
the feature vectors extracted by the convolutional neural network are feature vectors of size 1 x 100.
7. The method for identifying load of residents coupling neural network and dynamic time programming as claimed in claim 1, wherein said long-short term memory artificial neural network is structured to include: an input layer with dimension (1, 100), a BilSTM layer with dimension (1, 100), wherein the activation functions are sigmoid and tanh; a fully-connected layer with dimension (1, 11); a classification layer with dimension (1, 9), wherein the activation function is Softmax;
the training method of the long-short term memory artificial neural network comprises the following steps:
inputting the electric appliance characteristic vectors in the load database into a long-term and short-term memory artificial neural network for training;
in the training process, a gradient threshold value is set to be 1, a training period is set to be 60, each period is iterated for 54 times, and the training process uses an adaptive moment estimation optimization algorithm.
8. The method for identifying loads of residents coupled with neural network and dynamic time programming as claimed in claim 1, wherein the first set of probability vectors P is coupledBiLSTMWith a second set of probability vectors PDTWThe method for obtaining the final identification result comprises the following steps:
combining the probability vector sets of the two algorithms through the following formula to form a final identification vector set Pfinal＝{Pfinal(1),Pfinal(2),…,Pfinal(i)…,Pfinal(k)}：
Figure FDA0003269563970000031
In the formula, Pfinal(i) Representing the final probability value of the electric appliance to be identified as an electric appliance i, wherein the value range of i is 1 to k, k is the total number of the electric appliances which can be identified in the database, and P is the total number of the electric appliances which can be identified in the databaseBiLSTM＝{PBiLSTM(1),PBiLSTM(2),…,PBiLSTM(i)…,PBiLSTM(k) Represents the set of probability values, P, of the output of the BILSTM modelBiLSTM(i) Representing the probability value of the electric appliance i identified by the CNN-BilSTM network; pDTW＝{PDTW(1),PDTW(2),…,PDTW(i)…,PDTW(k) Represents the DTW probability vector set, P, of the DTW algorithm outputDTW(i) Representing the probability value of the electric appliance i identified by the DTW algorithm;
comparing the final set of discriminatory vectors PfinalThe maximum element in the set is selected according to the size of each element in the electric appliance to be identified, and then the final identification result of the electric appliance to be identified can be obtained.
9. A residential load identification device coupling a neural network with dynamic time planning, the device comprising:
an acquisition module: the method comprises the steps of obtaining a current and voltage time sequence of an electric appliance to be identified;
a steady state extraction module: the device is used for extracting the steady-state operation state of the electric appliance to be identified through the probability mass function;
a curve extraction module: extracting a current and voltage time sequence of the electric appliance to be identified in a steady state operation state, carrying out normalization processing to obtain a processed sequence, and drawing a V-I characteristic curve gray scale graph according to the processed sequence;
the vector extraction module: the system is used for inputting the V-I characteristic curve gray scale map into a convolution neural network and extracting a characteristic vector;
identifying the coupling module: inputting the characteristic vector into a trained long-short term memory artificial neural network for identification to obtain a first probability vector set PBiLSTM(ii) a Carrying out similarity comparison with the characteristic vectors in the load database by using a dynamic time normalization algorithm to obtain a second probability vector set PDTWCoupling a first set of probability vectors PBiLSTMWith a second set of probability vectors PDTWAnd obtaining a final identification result.
CN202111097344.5A 2021-09-18 2021-09-18 Method and device for identifying load of residents by coupling neural network and dynamic time planning Pending CN113987910A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111097344.5A CN113987910A (en) 2021-09-18 2021-09-18 Method and device for identifying load of residents by coupling neural network and dynamic time planning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111097344.5A CN113987910A (en) 2021-09-18 2021-09-18 Method and device for identifying load of residents by coupling neural network and dynamic time planning

Publications (1)

Publication Number Publication Date
CN113987910A true CN113987910A (en) 2022-01-28

Family

ID=79736116

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111097344.5A Pending CN113987910A (en) 2021-09-18 2021-09-18 Method and device for identifying load of residents by coupling neural network and dynamic time planning

Country Status (1)

Country Link
CN (1) CN113987910A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117333724A (en) * 2023-11-28 2024-01-02 天津滨电电力工程有限公司 Non-invasive load identification method based on multi-feature fusion image
CN117540326A (en) * 2024-01-09 2024-02-09 深圳大学 Construction state abnormality identification method and system for tunnel construction equipment by drilling and blasting method

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117333724A (en) * 2023-11-28 2024-01-02 天津滨电电力工程有限公司 Non-invasive load identification method based on multi-feature fusion image
CN117333724B (en) * 2023-11-28 2024-02-27 天津滨电电力工程有限公司 Non-invasive load identification method based on multi-feature fusion image
CN117540326A (en) * 2024-01-09 2024-02-09 深圳大学 Construction state abnormality identification method and system for tunnel construction equipment by drilling and blasting method
CN117540326B (en) * 2024-01-09 2024-04-12 深圳大学 Construction state abnormality identification method and system for tunnel construction equipment by drilling and blasting method

Similar Documents

Publication Publication Date Title
CN108573225B (en) Partial discharge signal pattern recognition method and system
CN111860982A (en) Wind power plant short-term wind power prediction method based on VMD-FCM-GRU
CN110111015A (en) A kind of power quality analysis method based on the multiple dimensioned arrangement entropy of variation mode decomposition
CN109147817B (en) Denoising frequency characteristic extraction method based on variation-limited Boltzmann machine
Wu et al. A hybrid support vector regression approach for rainfall forecasting using particle swarm optimization and projection pursuit technology
CN111008224B (en) Time sequence classification and retrieval method based on deep multitasking representation learning
CN113987910A (en) Method and device for identifying load of residents by coupling neural network and dynamic time planning
CN110880369A (en) Gas marker detection method based on radial basis function neural network and application
CN112381248A (en) Power distribution network fault diagnosis method based on deep feature clustering and LSTM
CN115659254A (en) Power quality disturbance analysis method for power distribution network with bimodal feature fusion
CN108919067A (en) A kind of recognition methods for GIS partial discharge mode
CN111859010A (en) Semi-supervised audio event identification method based on depth mutual information maximization
CN112364974B (en) YOLOv3 algorithm based on activation function improvement
CN112596016A (en) Transformer fault diagnosis method based on integration of multiple one-dimensional convolutional neural networks
CN111222689A (en) LSTM load prediction method, medium, and electronic device based on multi-scale temporal features
CN116187835A (en) Data-driven-based method and system for estimating theoretical line loss interval of transformer area
CN115100466A (en) Non-invasive load monitoring method, device and medium
CN117154256A (en) Electrochemical repair method for lithium battery
CN111090679B (en) Time sequence data representation learning method based on time sequence influence and graph embedding
CN116561569A (en) Industrial power load identification method based on EO feature selection and AdaBoost algorithm
CN116482491A (en) Transformer partial discharge fault diagnosis method based on Bayesian neural network
CN116186513A (en) Vibration signal identification method based on one-dimensional convolutional neural network
CN114841266A (en) Voltage sag identification method based on triple prototype network under small sample
CN113822771A (en) Low false detection rate electricity stealing detection method based on deep learning
CN114021424A (en) PCA-CNN-LVQ-based voltage sag source identification method

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