CN114722873A - Non-invasive load decomposition method based on residual convolution and attention mechanism - Google Patents

Non-invasive load decomposition method based on residual convolution and attention mechanism Download PDF

Info

Publication number
CN114722873A
CN114722873A CN202210393278.4A CN202210393278A CN114722873A CN 114722873 A CN114722873 A CN 114722873A CN 202210393278 A CN202210393278 A CN 202210393278A CN 114722873 A CN114722873 A CN 114722873A
Authority
CN
China
Prior art keywords
decomposition
residual
model
data
layer
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.)
Withdrawn
Application number
CN202210393278.4A
Other languages
Chinese (zh)
Inventor
易灵芝
许翔翔
刘江永
廖剑霞
刘西蒙
罗晓雪
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.)
Xiangtan University
Original Assignee
Xiangtan University
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 Xiangtan University filed Critical Xiangtan University
Priority to CN202210393278.4A priority Critical patent/CN114722873A/en
Publication of CN114722873A publication Critical patent/CN114722873A/en
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G06N5/041—Abduction
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S20/00—Management or operation of end-user stationary applications or the last stages of power distribution; Controlling, monitoring or operating thereof
    • Y04S20/20—End-user application control systems
    • Y04S20/242—Home appliances

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Signal Processing (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a non-invasive load decomposition method based on residual convolution and attention mechanism. Load characteristics are extracted through a residual convolution neural network, the thought of cross-layer connection is used for reference, the defects of gradient dispersion and performance degradation caused by network deepening are well overcome, information transmission before and after the network is smoother, and the problems that a network model has gradient disappearance and the like are avoided. And then, an attention mechanism is introduced, useless redundant data are removed, the data characterization capability is further enhanced, and a decomposition result is output through a full connection layer. And training a non-invasive load decomposition model by using a large amount of sample data, continuously fine-tuning model parameters, constructing an optimized decomposition model, and completing load decomposition by using the optimized non-invasive load decomposition model.

Description

Non-invasive load decomposition method based on residual convolution and attention mechanism
Technical Field
The invention relates to a non-intrusive load decomposition method based on residual convolution and an attention mechanism, which can be widely applied to non-intrusive load on-line monitoring of electrical loads of contract-signing users and residents in towns and rural areas, and has the characteristics of high accuracy, high decomposition speed and the like.
Background
The electric energy plays a vital role in the development of society, the demand of human beings on the electric energy is continuously increased, and the reasonable use of the electric energy has a great influence on the economic development of the whole society. Therefore, the utilization rate of the electric energy is effectively improved, the distribution of the electric energy resources is reasonably planned, and the urgent need for solving the problem of social sustainable development is met. Under the normal condition, according to an electric meter installed outdoors, a user only knows how much electricity is always shared, but cannot know the specific information of load energy consumption of each household appliance, and in order to optimally schedule adjustable loads such as an air conditioner, an electric automobile, a washing machine, a water heater, a disinfection cabinet, a printer and the like, under the condition that the electricity utilization comfort of the user is allowed, the highest utilization rate of new energy, the lowest total electricity fee based on time-of-use electricity price and the lowest total power peak-valley difference are realized, and the load monitoring of the user is needed.
With the development of smart grids, the traditional power industry is developing in a direction of high density, knowledge type and technology type, and besides the quality and quantity of the power generation side, the management of the demand side should be emphasized. Therefore, the demand for intelligence on the distribution and consumption sides is increasing. In order to enable an energy management part to effectively guide a signed user providing a main high-power household appliance electricity utilization habit, know the household charge energy consumption condition of the user and help the user to actively participate in demand response to the maximum extent, a technical key method is to realize the power load monitoring of the signed user. At present, two methods of invasive load monitoring and non-invasive load monitoring are mainly available. The intrusive load monitoring method includes the steps that each power load is required to be provided with a monitoring device for real-time monitoring, then load data are transmitted in a local storage or wireless communication mode, and the load data are analyzed on an upper computer to achieve load power utilization optimized scheduling; the monitoring mode has high data acquisition reduction degree and precision, but the installation and maintenance cost of the device is high; in addition, each load installation monitoring may cause a user privacy disclosure problem, and most users may generate a conflicting psychology. The non-invasive load monitoring method only needs to install a monitoring device (namely an existing intelligent ammeter) at the power supply inlet end of each user, collects power information such as total voltage, current, active power and reactive power of a power load in real time, and completes the decomposition of various loads by using a decomposition algorithm.
The hardware part of the non-invasive load monitoring and decomposing device is an intelligent measuring device at the client side. As shown in fig. 1, the device monitors the power consumption of various electric devices under the device, and actually, the NILMD function can be integrated into the smart meter. The electric quantities of the total load, such as voltage, current and the like, measured by the NILMD device can be regarded as signals carrying electric power information and comprise load component information with different characteristics, and by extracting the characteristic information of the electric quantities, the NILMD system can realize load decomposition and estimate the energy utilization information of the use state, the energy consumption and the like of the single electric equipment.
Disclosure of Invention
The invention mainly aims to provide a non-invasive resident load decomposition method based on residual convolution and an attention mechanism.
The method comprises the following steps:
step1, collecting power data by using a non-invasive load monitoring and decomposition (NILMD) measuring device, carrying out data preprocessing, and constructing a database by using a large amount of sample data;
step2 non-invasive load decomposition, extracting features of input data through a residual convolution neural network to obtain a feature map;
step3, an attention mechanism is introduced, the characteristic data are further processed, favorable characteristics are more effectively extracted, useless redundant characteristics are discarded, and the non-invasive load decomposition efficiency is improved;
step4, outputting a decomposition result through the full connection layer;
and Step5, repeating the steps of Step2-Step4, training the decomposition model by using the sample data, adjusting parameters and constructing a non-invasive load decomposition model. And then, taking the total power time sequence in the test set as the input of the load decomposition model to obtain a decomposition result, and analyzing the decomposition result.
The invention provides a non-invasive resident load decomposition method based on residual convolution and an attention mechanism. Load characteristics are extracted through a residual convolutional neural network, the thought of 'cross-layer connection' is used for reference, the defects of gradient dispersion and performance degradation caused by network deepening are well overcome, information is transmitted more smoothly before and after the network, and the problems that a network model has gradient disappearance and the like are avoided. And then, an attention mechanism is introduced, useless redundant data are eliminated, the data characterization capability is further enhanced, and a decomposition result is output through a full connection layer. Training a non-invasive load decomposition model by using a large amount of sample data, continuously fine-tuning model parameters, constructing an optimized decomposition model, and completing load decomposition by using the optimized non-invasive load decomposition model.
Drawings
In order to make the reader more clearly understand the embodiments of this patent, the following brief description of the drawings in the detailed description of this patent is provided:
FIG. 1 is a non-intrusive load resolution hardware implementation;
FIG. 2 is a non-intrusive load decomposition model;
FIG. 3 is a block diagram of a non-intrusive load decomposition residual convolutional neural network;
FIG. 4 is a non-intrusive load split channel attention mechanism module;
FIG. 5 is a non-intrusive load resolution spatial attention mechanism module;
FIG. 6 is a non-intrusive load resolution fully-connected layer output module;
Detailed Description
The invention provides a non-invasive resident load decomposition method based on residual convolution and an attention mechanism.
The method comprises the following steps:
step1, collecting power data by using a non-invasive load monitoring and decomposition (NILMD) measuring device, carrying out data preprocessing, and constructing a database by using a large amount of sample data;
step2 non-invasive load decomposition, extracting features of input data through a residual convolution neural network to obtain a feature map;
step3, an attention mechanism is introduced, the characteristic data are further processed, favorable characteristics are more effectively extracted, useless redundant characteristics are discarded, and the non-invasive load decomposition efficiency is improved;
step4, outputting a decomposition result through the full connection layer;
and Step5, repeating the steps of Step2-Step4, training the decomposition model by using the sample data, adjusting parameters and constructing a non-invasive load decomposition model. And then, taking the total power time sequence in the test set as the input of a non-invasive load decomposition model to obtain a decomposition result, and analyzing the non-invasive load decomposition result.
1. The method of claim 1, further comprising collecting power data using a non-invasive load monitoring and decomposition (NILMD) measurement device, and performing data preprocessing to construct the database using a plurality of sample data. The method is characterized in that: the data acquisition is the first step of non-invasive load decomposition, the data of the non-invasive load decomposition are acquired by the intelligent electric meter deployed in a specific area, and the intelligent electric meter integrates the function of a non-invasive load monitoring and measuring device to complete the data acquisition. Selecting a resident contracted user in a certain area as a research object, collecting a total power time sequence and a power time sequence of a single device, after all data are synthesized, carrying out normalization processing on the data, and finally dividing the data into a training set and a test set according to the proportion of 8:2 to complete the construction of a database.
2. The non-invasive load decomposition of claim 2 performs feature extraction on input data by a residual convolutional neural network. The method is characterized in that the number of layers of the neural network is continuously deepened along with the wide application of the neural network model in non-invasive load decomposition, and the problems of data loss, gradient disappearance and the like are inevitable. In order to make up the defects of insufficient residential load characteristic utilization rate, data loss, poor decomposition effect of low-utilization-rate electrical appliances, gradient disappearance and the like, a residual convolution neural network is provided, and the model has good feature extraction capability and learning capability and lays a foundation for improving the decomposition accuracy of a non-invasive load decomposition model.
The traditional convolutional neural network is insufficient in load feature extraction and also causes data loss, and in order to improve the feature extraction capability, a residual convolutional neural network is adopted to extract features of input data. Due to the deepening of the number of layers of the neural network model, when errors are reversely propagated to the front layer, parameter disturbance is basically difficult to generate, parameters of each layer cannot be updated, and the problem that the training data are difficult to learn by the network is caused. And the residual convolutional neural network can well make up the defects of performance degradation caused by gradient dispersion and network deepening. The residual convolution neural network draws the thought of 'cross-layer connection', adds the original input feature mapping and the output feature mapping of the rear layer through the connection mode, and completes the fusion of the features by utilizing the activation function activation, and the residual convolution neural network enables the front and rear information transmission of the network to be smoother. As shown in fig. 1, the residual structure is defined by defining the bottom layer mapping as h (x), and using the stack layer to fit f (x) ═ h (x) + x, and adding the input x of the residual unit, the output of the residual unit is obtained. F (x) is the residual function. In the basic structure of residual concatenation, the residual concatenation is mainly composed of two weight layers, each layer containing weight parameters. x represents the input and the residual join expression is as follows:
F(x)＝W2σ(W1x) (1)
where σ denotes the nonlinear activation function of ReLU, W1,W2Is a weight parameter. And finally, adding the input end x and the output result of the second weight layer in a cross-layer connection transfer mode to obtain the final output of the whole residual error structure. The calculation formula is as follows:
y＝F(x,{Wi})+x (2)
wherein, F (x, { W)i}) is the residual mapping function to be trained and learned in residual concatenation, { WiIs the set of weight parameters. According to the formulas (1) and (2), it can be obtained:
Figure BDA0003596189180000031
wherein x isLIs the output characteristic of the L-th residual connection in the residual structureMapping, xiIs the input feature map of the 1 st residual connection in the residual structure. It can be concluded that the output feature map of the lth residual unit can be decomposed as the sum of the input feature map of the ith residual unit and the feature maps between the two residual connections. The performance of feature extraction can be improved through the residual convolution neural network, and the problem of gradient disappearance caused by the increase of the number of layers of the neural network can be avoided. The number of convolution kernels is 40, and the size is 3 x 3. The step size is 1 and the ReLU function is chosen as the activation function. The residual convolutional neural network module is shown in figure 3.
3. An attention-calling mechanism, as recited in claim 3, further processes the feature data. The method is characterized in that: with the rapid development of the deep learning algorithm, more and more models are applied to the task of non-intrusive load decomposition, and good performance is mostly obtained. The establishment of the variant models is based on different networks, and the efficiency of the models is improved by combining and optimizing. However, as the model structure continuously deepens, the parameter quantity is excessive, the training speed is slow, and the calculation cost is too high, so that the phenomena of feature redundancy and the decline of the learning capability of certain shallow networks can occur. The expression of the features is enhanced from two dimensions of a channel and a space by introducing an attention mechanism, so that favorable features can be effectively extracted, useless features are discarded, the training time is effectively reduced, and the accuracy of non-invasive load decomposition is obviously improved.
Firstly, carrying out feature recalibration on input multi-channel features by using a channel attention module; then, carrying out re-weighting integration on different position information of the input features by using a space attention module to enhance the interested region; and finally, fusing the enhanced information to obtain an enhanced characteristic result. As shown in fig. 4, the H × W × C data feature is obtained through the residual convolutional neural network, and then, through the channel attention mechanism, in order to obtain the weight vector of the channel, two 1 × 1 convolutional layer compression channels are adopted for performing the global average pooling operation on the input at the beginning; then restoring the channels to fuse the information between the channels; finally, the 1 × 1 × C vector is converted into a weight vector by a Sigmoid function. This process can be described as:
Wca＝f{WU{δ{WD[G(z)]}}} (4)
where z is the input to the channel attention module, G denotes global average pooling, WDAnd WURepresents the 1 x 1 convolutional layers for channel down and up dimensions, δ and f represent the corrected linear unit (ReLU) and Sigmoid activation functions, respectively. And after the feature recalibration is carried out through the channel attention mechanism module, the information is aggregated from different positions in the feature mapping after entering the space attention mechanism module. The spatial attention mechanism module is shown in fig. 5, the input of the module is a H × W feature map z with the number of channels being C, firstly, global average pooling is performed on low-level features in the channel direction to reduce dimension, a feature map with the size being H × W is obtained, certain weight is learned through convolution, secondly, a single-channel feature map is changed into H × W one-dimensional features through deformation operation, global information of each feature position relative to the whole feature map is obtained through Softmax, and then the one-dimensional features containing the global information are changed into H × W feature map M.
M＝Softmax[G(z)] (5)
Finally, M passes through a convolution layer and a Sigmoid activation function to generate a space attention weight Wsa. The following formula:
Wsa＝Sigmoid[Conv(M)] (6)
in the formula: conv denotes the convolution operation. The final weighted feature output is:
Figure BDA0003596189180000041
in the formula
Figure BDA0003596189180000042
Representing element multiplication. And carrying out weighted assignment on the features containing the channel information by using a spatial attention module to acquire global information of the features, and giving a spatial attention weight so that the network learns the feature information useful for the decomposition task according to weight distribution. Universal jointThrough the double attention module, the network can fully utilize information of different channels and positions in the characteristic diagram, important information is gathered to weaken useless information, attention of the network to key characteristics is improved, and more effective characteristic data is obtained.
4, as claimed in claim 4, outputting the decomposition result through the full connection layer. The method is characterized in that: as shown in fig. 6, the non-intrusive load decomposition model inputs the feature data passing through the dual-attention machine model into the fully-connected layer, uses the combination of the fully-connected layer to implement the non-linear mapping from the deep features to the target electrical appliance power, and combines with the seq2point learning model to obtain a decomposition result. The fully-connected layer has three layers, the dimensionalities are respectively 64, 32 and 16, and the last fully-connected layer and the output form a seq2point model. The calculation for each fully connected layer is as follows:
Figure BDA0003596189180000051
wherein the content of the first and second substances,
Figure BDA0003596189180000052
is the input of the previous layer, and the input of the next layer,
Figure BDA0003596189180000053
is the output of this layer.
The principle of the Seq2point model is that the midpoint of a window is predicted by training a network instead of the data of the whole window, and the window data input by the network is assumed to be Xt:t+W-1W represents the window length and the output is the midpoint y of the corresponding window of the target deviceτWhere the point element is a non-linear function of the power window, t + W/2, and Seq2point defines a neural network that will input a sliding window Xt:t+W-1Mapping to a corresponding window Yt:t+W-1The midpoint of (a).
yτ＝f(Xt:t+W-1)+ε (9)
To handle the end points of the sequence, at the input power sequence X ═ X1,x2,…,xT) The first end and the last end are filled with W/2 zero elements, and the model has the advantages that each zero element isx corresponds to one prediction instead of the average prediction per window.
To prevent overfitting of the model, add the L2 regularized mean square error loss function:
Figure BDA0003596189180000054
in the formula: y istThe real value of the model;
Figure BDA0003596189180000055
is a model predicted value; λ is the weight attenuation coefficient.
5. The method of claim 5, wherein the steps of Step2-Step4 are repeated, the decomposition model is trained by using the sample data, the parameters are adjusted, and the non-invasive load decomposition model is constructed. And then, taking the total power time sequence in the test set as the input of the load decomposition model to obtain a decomposition result, and analyzing the decomposition result. The method is characterized in that: as shown in fig. 2, the whole non-invasive load decomposition model is trained through a sample set, training samples are divided into a plurality of batches and sent into a network according to the batches for training, Batchsize is set to be 1000, the training times are 50, an Adam optimizer is used for back propagation to correct network parameters until Loss converges, and meanwhile, an early suppression mechanism is added, namely after the mean square error stops decreasing and reaches a certain iteration times, the training is forcibly ended, and the prediction classification accuracy of the model can be improved by optimizing the model parameters. When the loss function converges, the model training is completed, and the learning rate is set to 0.001, which has the best effect. And then inputting the test set data into the trained decomposition model to obtain the decomposition result of the signed user, namely the power sequence of single loads such as a refrigerator, a washing machine, an air conditioner, a lamp, a computer and the like.
To evaluate the performance of the load decomposition, the mean absolute error MAE, the normalized signal total error SAE and the normalized decomposition error NDE were used for evaluation herein, respectively
Figure BDA0003596189180000056
Figure BDA0003596189180000057
Figure BDA0003596189180000061
In the formula: x is the number oftRepresenting the true value of the device at time t;
Figure BDA0003596189180000062
representing a predicted value of the equipment at time t; r ═ ΣrxtRepresenting the actual total energy consumption of the plant;
Figure BDA0003596189180000063
indicating that the plant is predicting the total energy consumption. By verification, the feasibility of the decomposition model is proved.

Claims (6)

1. A non-invasive load decomposition method based on residual convolution and an attention mechanism is characterized by comprising the following steps:
step1, collecting power data by using a non-invasive load monitoring and decomposition (NILMD) measuring device, carrying out data preprocessing, and constructing a database by using a large amount of sample data;
step2 non-invasive load decomposition, extracting features of input data through a residual convolution neural network to obtain a feature map;
step3, an attention mechanism is introduced, the characteristic data are further processed, favorable characteristics are more effectively extracted, useless redundant characteristics are discarded, and the non-invasive load decomposition efficiency is improved;
step4, outputting a decomposition result through the full connection layer;
and Step5, repeating the steps of Step2-Step4, training the decomposition model by using the sample data, adjusting parameters and constructing a non-invasive load decomposition model. And then, taking the total power time sequence in the test set as the input of a non-invasive load decomposition model to obtain a decomposition result, and analyzing the non-invasive load decomposition result.
2. The method of claim 1, wherein the power data is collected and pre-processed by a non-invasive load monitoring and decomposition (NILMD) measurement device, and a database is constructed by using a large amount of sample data; the method is characterized in that: the method comprises the following steps of collecting data, namely, collecting the data of non-invasive load decomposition by an intelligent electric meter deployed in a specific area, wherein the intelligent electric meter integrates the function of a non-invasive load monitoring and measuring device and can finish data collection; selecting a resident contracted user in a certain area as a research object, collecting a total power time sequence and a power time sequence of a single device, after all data are synthesized, carrying out normalization processing on the data, and finally dividing the data into a training set and a test set according to the proportion of 8:2 to complete the construction of a database.
3. The non-invasive load decomposition of claim 2, performing feature extraction on input data through a residual convolutional neural network; the method is characterized in that as the neural network model is widely applied to non-invasive load decomposition, the number of neural network layers is continuously deepened, so that the problems of data loss, gradient disappearance and the like are inevitable; in order to make up the defects of insufficient residential load characteristic utilization rate, data loss, poor decomposition effect of low-utilization-rate electrical appliances, gradient disappearance and the like, a residual convolution neural network is provided, the model has good feature extraction capability and learning capability, and a foundation is laid for improving the decomposition accuracy of a non-invasive load decomposition model;
the traditional convolutional neural network is insufficient in load feature extraction and also causes data loss, and in order to improve the feature extraction capability, a residual convolutional neural network is adopted to extract features of input data; due to the deepening of the number of layers of the neural network model, when errors are reversely propagated to the front layer, parameter disturbance is basically difficult to generate, so that parameters of each layer cannot be updated, and the problem that the training data are difficult to learn by the network is caused; the residual convolutional neural network can well make up the defects of performance degradation caused by gradient dispersion and network deepening; the residual convolutional neural network uses the thought of 'cross-layer connection', adds the original input feature mapping and the output feature mapping of the rear layer through the connection mode, and completes the fusion of the features by utilizing the activation function activation, and the residual convolutional neural network enables the front and rear information transmission of the network to be smoother; as shown in fig. 1, the residual structure defines a bottom layer mapping as h (x), and the stack layer is used to fit f (x) ═ h (x) + x, and then the input x of the residual unit is added to obtain the output of the residual unit; f (x) is the residual function. In the basic structure of residual concatenation, the residual concatenation is mainly composed of two weight layers, each layer containing weight parameters. x represents the input and the residual join expression is as follows:
F(x)＝W2σ(W1x) (1)
where σ denotes the nonlinear activation function of ReLU, W1,W2Is a weight parameter. Finally, adding the input end x and the output result of the second weight layer in a cross-layer connection transfer mode to obtain the final output of the whole residual error structure; the calculation formula is as follows:
y＝F(x,{Wi})+x (2)
wherein, F (x, { W)i}) is the residual mapping function to be trained and learned in residual concatenation, { WiIs the set of weight parameters; according to the formulas (1) and (2), it can be obtained:
Figure FDA0003596189170000021
wherein x isLIs the output feature map, x, of the L-th residual connection in the residual structureiIs the input feature mapping of the 1 st residual connection in the residual structure; it can be concluded therefrom that the output feature map of the lth residual unit can be decomposed into the sum of the input feature map of the ith residual unit and the feature maps between the two residual connections; the performance of feature extraction can be improved through the residual convolution neural network, and the problem of gradient disappearance caused by the increase of the layer number of the neural network can be avoided; the number of convolution kernels is 40, and the size is 3 multiplied by 3. The step length is 1, and a ReLU function is selected as an activation function; residual convolutional neural networkThe module is shown in figure 3.
4. An attention-calling mechanism, as recited in claim 3, further processing the feature data; the method is characterized in that: with the rapid development of a deep learning algorithm, more and more models are applied to a task of non-intrusive load decomposition, and good performance is mostly obtained; the establishment of the variant models is based on different networks, and the efficiency of the models is improved by combination and optimization; however, as the model structure continuously deepens, the parameter quantity is excessive, the training speed is slow, and the calculation cost is too high, so that the phenomena of feature redundancy and the decline of learning capacity of certain shallow networks can occur; the expression of the features is enhanced from two dimensions of a channel and a space by introducing an attention mechanism, so that favorable features can be effectively extracted, useless features are discarded, the training time is effectively reduced, and the accuracy of non-invasive load decomposition is obviously improved;
firstly, carrying out feature recalibration on input multi-channel features by using a channel attention module; then, carrying out re-weighting integration on different position information of the input features by using a space attention module to enhance the interested region; finally, fusing the enhanced information to obtain an enhanced characteristic result; as shown in fig. 4, the H × W × C data feature is obtained through the residual convolutional neural network, and then, through the channel attention mechanism, in order to obtain the weight vector of the channel, two 1 × 1 convolutional layer compression channels are adopted for performing the global average pooling operation on the input at the beginning; then restoring the channels to fuse the information between the channels; finally, converting the 1 multiplied by C vector into a weight vector through a Sigmoid function; this process can be described as:
Wca＝f{WU{δ{WD[G(z)]}}} (4)
where z is the input to the channel attention module, G denotes global average pooling, WDAnd WURepresents the 1 × 1 convolutional layer for channel dimensionality reduction and dimensionality enhancement, δ and f represent the corrected linear unit (ReLU) and Sigmoid activation functions, respectively; after the characteristic recalibration is carried out through the channel attention mechanism module, the channel attention mechanism module enters a space attention mechanism moduleAggregating information from different locations in the feature map; as shown in fig. 5, the spatial attention mechanism module inputs an H × W feature map z with a channel number of C, performs global average pooling on low-level features in a channel direction to reduce dimension, obtains a H × W feature map, learns a certain weight through convolution, changes a single-channel feature map into H × W one-dimensional features through a morphing operation, obtains global information of each feature position relative to the entire feature map through Softmax, and then changes the one-dimensional features containing the global information into a H × W feature map M:
M＝Soft max[G(z)] (5)
finally, M passes through a convolution layer and a Sigmoid activation function to generate a space attention weight Wsa(ii) a The following formula:
Wsa＝Sigmoid[Conv(M)] (6)
in the formula: conv denotes the convolution operation. The final weighted feature output is:
Figure FDA0003596189170000031
in the formula
Figure FDA0003596189170000032
Representing element multiplication; carrying out weighted assignment on the characteristics containing the channel information by using a space attention module to obtain global information of the characteristics, and giving space attention weight so that the network learns the characteristic information useful for decomposing tasks according to weight distribution; through the double attention modules, the network can fully utilize information of different channels and positions in the characteristic diagram, important information is gathered to weaken useless information, the attention of the network to key characteristics is improved, and more effective characteristic data is obtained.
5. Outputting the decomposition result through the full connection layer as claimed in claim 4; the method is characterized in that: as shown in fig. 6, the non-intrusive load decomposition model inputs the feature data passing through the dual-attention machine model into the full connection layer, uses the combination of the full connection layer to realize the non-linear mapping from the deep features to the power of the target electrical appliance, and combines with the seq2point learning model to obtain a decomposition result; the three full-link layers are three, the dimensionalities are 64, 32 and 16 respectively, and the last full-link layer and the output form a seq2point model. The calculation for each fully connected layer is as follows:
Figure FDA0003596189170000033
wherein, the first and the second end of the pipe are connected with each other,
Figure FDA0003596189170000034
is the input of the previous layer, and the input of the next layer,
Figure FDA0003596189170000035
is the output of this layer;
the principle of the Seq2point model is that the midpoint of a window is predicted by training a network instead of the data of the whole window, and the window data input by the network is assumed to be Xt:t+W-1W represents the window length and the output is the midpoint y of the corresponding window of the target deviceτWhere the point element is a non-linear function of the power window, t + W/2, and Seq2point defines a neural network that will input a sliding window Xt:t+W-1Mapping to a corresponding window Yt:t+W-1A midpoint of (a);
yτ＝f(Xt:t+W-1)+ε (9)
to handle the end points of the sequence, at the input power sequence X ═ X1,x2,…,xT) W/2 zero elements are filled at the first end and the last end, and the model has the advantages that each x corresponds to one prediction instead of the average prediction of each window;
to prevent overfitting of the model, add the L2 regularized mean square error loss function:
Figure FDA0003596189170000036
in the formula：ytThe real value of the model;
Figure FDA0003596189170000037
is a model predicted value; λ is the weight attenuation coefficient.
6. The method of claim 5, wherein the steps of Step2-Step4 are repeated, the decomposition model is trained by using the sample data, parameters are adjusted, and a non-invasive load decomposition model is constructed; then, the total power time sequence in the test set is used as the input of a load decomposition model to obtain a decomposition result, and the decomposition result is analyzed; the method is characterized in that: as shown in fig. 2, training the whole non-intrusive load decomposition model through a sample set, dividing training samples into a plurality of batches, sending the batches into a network for training, setting Batchsize to 1000, training the number of times to be 50, using an Adam optimizer to reversely propagate and correct network parameters until Loss converges, and simultaneously adding an early timing mechanism, namely forcibly ending training after a mean square error stops decreasing to reach a certain number of iterations, and optimizing model parameters to improve the prediction classification accuracy of the model; when the loss function is converged, the model training is finished, and the learning rate is set to be 0.001, so that the best effect is achieved; then inputting the test set data into the trained decomposition model to obtain the decomposition result of the signed user, namely the power sequence of single loads such as a refrigerator, a washing machine, an air conditioner, a lamp and a computer;
to evaluate the performance of the load decomposition, the mean absolute error MAE, the normalized signal total error SAE and the normalized decomposition error NDE were used for evaluation herein, respectively
Figure FDA0003596189170000041
Figure FDA0003596189170000042
Figure FDA0003596189170000043
In the formula: x is the number oftRepresenting the true value of the device at time t;
Figure FDA0003596189170000044
representing a predicted value of the equipment at time t; r ═ ΣrxtRepresenting the actual total energy consumption of the plant;
Figure FDA0003596189170000045
indicating that the plant is predicting the total energy consumption. By verification, the feasibility of the decomposition model is proved.
CN202210393278.4A 2022-04-14 2022-04-14 Non-invasive load decomposition method based on residual convolution and attention mechanism Withdrawn CN114722873A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210393278.4A CN114722873A (en) 2022-04-14 2022-04-14 Non-invasive load decomposition method based on residual convolution and attention mechanism

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210393278.4A CN114722873A (en) 2022-04-14 2022-04-14 Non-invasive load decomposition method based on residual convolution and attention mechanism

Publications (1)

Publication Number Publication Date
CN114722873A true CN114722873A (en) 2022-07-08

Family

ID=82244618

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210393278.4A Withdrawn CN114722873A (en) 2022-04-14 2022-04-14 Non-invasive load decomposition method based on residual convolution and attention mechanism

Country Status (1)

Country Link
CN (1) CN114722873A (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115204529A (en) * 2022-09-15 2022-10-18 之江实验室 Non-invasive load monitoring method and device based on time attention mechanism
CN115201615A (en) * 2022-09-15 2022-10-18 之江实验室 Non-invasive load monitoring method and device based on physical constraint neural network
CN116522151A (en) * 2023-07-04 2023-08-01 深圳市芝电能源科技有限公司 Semi-supervised load recognition and decomposition method, system and storage medium
CN116611349A (en) * 2023-07-18 2023-08-18 华东交通大学 Neural network-based roller wire drying process parameter optimization method and system
CN116632842A (en) * 2023-07-26 2023-08-22 国网山东省电力公司信息通信公司 Clustering characteristic-based method and system for predicting distribution type photovoltaic load probability of platform
CN117808175A (en) * 2024-03-01 2024-04-02 南京信息工程大学 Short-term multi-energy load prediction method based on DTformer

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115204529A (en) * 2022-09-15 2022-10-18 之江实验室 Non-invasive load monitoring method and device based on time attention mechanism
CN115201615A (en) * 2022-09-15 2022-10-18 之江实验室 Non-invasive load monitoring method and device based on physical constraint neural network
CN115204529B (en) * 2022-09-15 2022-12-20 之江实验室 Non-invasive load monitoring method and device based on time attention mechanism
CN115201615B (en) * 2022-09-15 2022-12-20 之江实验室 Non-invasive load monitoring method and device based on physical constraint neural network
CN116522151A (en) * 2023-07-04 2023-08-01 深圳市芝电能源科技有限公司 Semi-supervised load recognition and decomposition method, system and storage medium
CN116611349A (en) * 2023-07-18 2023-08-18 华东交通大学 Neural network-based roller wire drying process parameter optimization method and system
CN116611349B (en) * 2023-07-18 2023-10-10 华东交通大学 Neural network-based roller wire drying process parameter optimization method and system
CN116632842A (en) * 2023-07-26 2023-08-22 国网山东省电力公司信息通信公司 Clustering characteristic-based method and system for predicting distribution type photovoltaic load probability of platform
CN116632842B (en) * 2023-07-26 2023-11-10 国网山东省电力公司信息通信公司 Clustering characteristic-based method and system for predicting distribution type photovoltaic load probability of platform
CN117808175A (en) * 2024-03-01 2024-04-02 南京信息工程大学 Short-term multi-energy load prediction method based on DTformer
CN117808175B (en) * 2024-03-01 2024-05-17 南京信息工程大学 DTformer-based short-term multi-energy load prediction method

Similar Documents

Publication Publication Date Title
CN114722873A (en) Non-invasive load decomposition method based on residual convolution and attention mechanism
Li et al. Probabilistic charging power forecast of EVCS: Reinforcement learning assisted deep learning approach
CN109546659B (en) Power distribution network reactive power optimization method based on random matrix and intelligent scene matching
CN106295899B (en) Wind power probability density Forecasting Methodology based on genetic algorithm Yu supporting vector quantile estimate
CN113112077B (en) HVAC control system based on multi-step prediction deep reinforcement learning algorithm
CN110443417A (en) Multi-model integrated load prediction method based on wavelet transformation
CN112052913B (en) Distributed photovoltaic power station power data virtual acquisition method
CN112215428B (en) Photovoltaic power generation power prediction method and system based on error correction and fuzzy logic
Aliberti et al. Forecasting Short-term Solar Radiation for Photovoltaic Energy Predictions.
CN115759371A (en) GCN-LSTM-based short-term load prediction method for power system
CN114418174A (en) Electric vehicle charging load prediction method
CN116629417A (en) Rail transit traction load prediction method combined with convolution attention module
CN115828755A (en) Method and device for evaluating participation of micro-grid group in power grid service and readable storage medium
CN115374908A (en) Short-term load prediction method and system suitable for flexible interconnection real-time control of distribution network
CN115759458A (en) Load prediction method based on comprehensive energy data processing and multi-task deep learning
CN113902183B (en) BERT-based non-invasive transformer area charging pile state monitoring and electricity price adjusting method
CN117540859B (en) Method and system for estimating generated energy of photovoltaic power station based on neural network
CN116756575B (en) Non-invasive load decomposition method based on BGAIN-DD network
Zhang et al. Hybrid model for renewable energy and load forecasting based on data mining and EWT
CN113705888A (en) Industrial steam generation amount prediction method and system based on Pearson correlation and neural network
CN111210051B (en) User electricity consumption behavior prediction method and system
Al-Hawani et al. Short-term forecasting of electricity consumption using gaussian processes
CN116167465A (en) Solar irradiance prediction method based on multivariate time series ensemble learning
CN113673846B (en) Demand situation awareness method under electric power Internet of things architecture
CN116523681A (en) Load decomposition method and device for electric automobile, electronic equipment and storage medium

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
WW01 Invention patent application withdrawn after publication
WW01 Invention patent application withdrawn after publication

Application publication date: 20220708