CN114745299B - Non-invasive load monitoring method based on sequence delay reconstruction CSP convolutional neural network - Google Patents
Non-invasive load monitoring method based on sequence delay reconstruction CSP convolutional neural network Download PDFInfo
- Publication number
- CN114745299B CN114745299B CN202210259630.5A CN202210259630A CN114745299B CN 114745299 B CN114745299 B CN 114745299B CN 202210259630 A CN202210259630 A CN 202210259630A CN 114745299 B CN114745299 B CN 114745299B
- Authority
- CN
- China
- Prior art keywords
- data
- matrix
- sequence
- csp
- convolutional neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L43/00—Arrangements for monitoring or testing data switching networks
- H04L43/08—Monitoring or testing based on specific metrics, e.g. QoS, energy consumption or environmental parameters
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L43/00—Arrangements for monitoring or testing data switching networks
- H04L43/04—Processing captured monitoring data, e.g. for logfile generation
-
- 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
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Signal Processing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Environmental & Geological Engineering (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Complex Calculations (AREA)
Abstract
The invention discloses a non-invasive load monitoring method based on a sequence delay reconstruction CSP convolutional neural network, which comprises the steps of firstly reconstructing a 2xL length time sequence in a data set into an LxLx 2-dimensional matrix M through sequence delay by utilizing the characteristic of array signal processing, constructing the convolutional neural network in a targeted manner by utilizing the characteristic of the CSP network according to the characteristic of the matrix M, training by utilizing the reconstructed matrix M data and the CSP convolutional neural network to obtain optimal network parameters, arranging the network at a monitoring end, loading corresponding parameters, and decomposing input data to obtain the working state of a load electric appliance; the invention realizes that the time sequence relation and the state change in the data are considered during monitoring, ensures the monitoring efficiency while improving the monitoring precision, can pertinently guide users or enterprises to reasonably and safely use electricity, assists an electric power supply department to perfect electric power dispatching work, and has the advantages of scientific and reasonable method, strong applicability, good effect and the like.
Description
Technical Field
The invention relates to the technical field of non-invasive load detection, in particular to a non-invasive load monitoring method based on a sequence delay reconstruction CSP convolutional neural network.
Background
Non-invasive load monitoring (Non-intrusive Load Monitoring, NILM) technology is capable of acquiring electricity utilization data of a user or an enterprise by installing monitoring equipment at a power inlet, and directly analyzing the data to obtain the type and the operation condition of a single load in a load cluster. By the technology, a user or an enterprise can obtain the working state of each load without integrating a large number of precise and expensive hardware sensor devices which are required to be connected into an original system circuit or professional monitoring personnel and special monitoring devices, and the technology is a nondestructive and practical load monitoring technology. With the development of intelligent electric meters, artificial intelligence and big data technologies in recent years, non-invasive load monitoring technologies have been greatly developed.
In recent years, after the deep learning algorithm is applied to non-invasive load monitoring, good performance is obtained, and the traditional non-invasive load monitoring algorithms such as a hidden Markov model, a combined optimization model and the like are gradually replaced. Wherein, whether the sequence-to-point algorithm or the sequence-to-sequence algorithm, higher accuracy and estimation precision can be obtained compared with the traditional algorithm. However, most of the researches only aim at the convolution characteristic extraction of the sequence data, so that the time sequence and state change in the sequence are not well applied and reflected; the correlation among QKV matrixes needs to be calculated in a large quantity according to various attention methods, the network structure is complex, the training cost is high, and engineering realization is difficult; therefore, there is a need to design a non-invasive load monitoring method based on the sequence delay reconstruction CSP (Cross Stage Partial) convolutional neural network.
Disclosure of Invention
The invention aims to overcome the defects of the prior art, and provides a non-invasive load monitoring method based on a sequence delay reconstruction CSP convolutional neural network, which aims to better solve the problems that most of the method only aims at carrying out convolutional feature extraction on sequence data, so that the time sequence relation and state change in the sequence are not well applied and reflected, a large amount of correlations among QKV matrixes are required to be calculated by a method of converting attention of a former class, the network structure is complex, the training cost is high, and engineering realization is difficult.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
a non-invasive load monitoring method based on a sequence delay reconstruction CSP convolutional neural network, which comprises the following steps,
utilizing the characteristic of array signal processing, reconstructing a 2xL length time sequence in a data set into a matrix M with LxLx2 dimension through sequence delay;
aiming at the characteristics of the matrix M, constructing a convolutional neural network in a targeted manner by utilizing CSP network characteristics;
training by using the reconstructed matrix M data and the CSP convolutional neural network to obtain optimal network parameters;
and (D) arranging a network at the monitoring end, loading corresponding parameters, and decomposing input data to obtain the working state of the load electric appliance.
The non-invasive load monitoring method based on the sequence delay reconstruction CSP convolutional neural network, step (A), the 2xL length time sequence in the data set is reconstructed into a matrix M of LxLx2 dimension through the sequence delay by utilizing the characteristic of array signal processing, the specific steps are as follows,
step (A1), a sequence with the length of 2xL is cut off into a sequence with the length of L by a time unit of sequence delay, and the total L+1 sequences are obtained;
reconstructing the first L sequences into an LxL dimension matrix A, reconstructing the last L sequences into an LxL dimension matrix B, subtracting the matrix A from the matrix B to obtain an LxL dimension matrix C, and splicing the matrix C and the matrix A into A2 xLxL matrix M; wherein the specific steps of the construction of the matrix M are as follows,
step (A22), the sequence of 2xL length is set as [ a ] 1 a 2 … a 2*L ]Then the LxL dimensional matrix A, B and C are obtained through transformation, and as shown in a formula (1),
wherein a is i Representing elements in the sequence, the subscripts of which represent the position of the data in the original sequence, i.e. a i For the ith sequence thereinThe first element of the column, a i+L-1 Is the last element in the original sequence;
step (A23), the matrix M is A2 xLxL-dimensional matrix formed by splicing an LxL-dimensional matrix C and an LxL-dimensional matrix A, as shown in a formula (2),
the non-invasive load monitoring method based on the sequence delay reconstruction CSP convolutional neural network comprises the following steps of (B) constructing the convolutional neural network by utilizing the CSP network characteristic according to the characteristic of the matrix M,
step (B1), after data is input into the network, firstly converting the data into matrix data through a sequence delay reconstruction module, and then completing downsampling through a convolution combination module Conv by using a convolution kernel with the size of 6x6 and the step length of 2, wherein the length and the width of the matrix are reduced to be 1/2 of the original length and the width of the matrix; the convolution combination module Conv consists of a standard two-dimensional convolution module Conv2d, a normalization module BatchNorm2d and an activation function module SiLU;
step (B2), sequentially passing through a plurality of groups of cross convolution splicing combination modules CSPcat and Conv modules, wherein the specific passing steps are as follows,
step (B21), when the data passes through the splicing and combining module CSPcat, two groups of data are obtained by two groups of modules Conv with the size of 1x1, the step length of 1 convolution kernel and the output channel number halved, wherein one group of data sequentially passes through the modules Conv with the size of 1x1, the step length of 1 and the output channel number halved, the modules Conv with the step length of 1 and the output channel number doubled, and the other group of data does not do any processing and do splicing operation;
step (B22), the data continue to pass through a normalization module BatchNorm2d, an activation function module SiLU and a module Conv with the size of 1x1 and the step length of 1 to obtain final output data, wherein the dimension of the output data is the same as that of the input data, an activation function formula used by the activation function module SiLU is shown as a formula (3),
f(x)＝x*sigmoid(x)＝x/(1+e -x )(3)。
the non-invasive load monitoring method based on the sequence delay reconstruction CSP convolutional neural network comprises the following steps of (C) training by using the reconstructed matrix M data and the CSP convolutional neural network to obtain optimal network parameters, wherein the training comprises the specific steps of converting the matrix M data into Linear layers and dimensional compression functions by a CSP convolutional neural network internal convolutional combination module Conv, a splicing combination module CSPcat, a flattening layer flat, linear layers and dimensional compression functions sequentially.
The non-invasive load monitoring method based on the sequence delay reconstruction CSP convolutional neural network comprises the following steps of (D) arranging the network at a monitoring end, loading corresponding parameters, and decomposing input data to obtain the working state of the load electric appliance, wherein the data dimension required by specific application is adapted by modifying the parameters of the last layers of the network.
The beneficial effects of the invention are as follows: the non-invasive load monitoring method based on the sequence delay reconstruction CSP convolutional neural network strengthens the influence of time sequence change in data on a monitoring result, simultaneously utilizes methods such as feature diagram splitting, residual network connection and the like, reduces the parameter quantity of the network, improves the accuracy of the network, effectively realizes the consideration of time sequence relation and state change in the data during monitoring, improves the monitoring precision and ensures the monitoring efficiency, so that a user or enterprise electricity data set is obtained at an electric power inlet through low-cost electric power monitoring equipment such as an intelligent ammeter and the like, and the designed network is trained by using the data set, and the type and state of the actually used electricity load of the user or enterprise can be obtained only by obtaining the low-frequency electricity data of the user or enterprise, thereby guiding the user or enterprise to reasonably and safely use electricity in a targeted manner, assisting an electric power supply department to perfect electric power dispatching work, and the like.
Drawings
FIG. 1 is a schematic overall flow diagram of a non-invasive load monitoring method based on a sequence delay reconstruction CSP convolutional neural network;
FIG. 2 is a schematic diagram of the network structure of the present invention with an input data dimension of 1 channel and a series length of 127;
FIG. 3 is a schematic view of the Conv module architecture of the present invention;
fig. 4 is a schematic diagram of the structure of the CSPcat module of the invention.
Detailed Description
The invention will be further described with reference to the drawings.
As shown in fig. 1-4, the non-invasive load monitoring method based on the sequence delay reconstruction CSP convolutional neural network of the invention includes the following steps,
a step (A) of reconstructing a 2xL length time sequence in a data set into a matrix M in LxLx2 dimension through sequence delay by utilizing the characteristic of array signal processing, which comprises the following specific steps,
wherein, the matrix M is added with target change information while maintaining time sequence information;
step (A1), a sequence with the length of 2xL is cut off into a sequence with the length of L by a time unit of sequence delay, and the total L+1 sequences are obtained;
reconstructing the first L sequences into an LxL dimension matrix A, reconstructing the last L sequences into an LxL dimension matrix B, subtracting the matrix A from the matrix B to obtain an LxL dimension matrix C, and splicing the matrix C and the matrix A into A2 xLxL matrix M; wherein the specific steps of the construction of the matrix M are as follows,
step (A22), the sequence of 2xL length is set as [ a ] 1 a 2 … a 2*L ]Then the LxL dimensional matrix A, B and C are obtained through transformation, and as shown in a formula (1),
wherein a is i Representing elements in the sequence, the subscripts of which represent the position of the data in the original sequence, i.e. a i A being the first element of the ith sequence, a i+L-1 Is the last element in the original sequence;
step (A23), the matrix M is A2 xLxL-dimensional matrix formed by splicing an LxL-dimensional matrix C and an LxL-dimensional matrix A, as shown in a formula (2),
aiming at the characteristics of the matrix M, the specific steps of constructing the convolutional neural network by utilizing the CSP network characteristics are as follows,
step (B1), after data is input into the network, firstly converting the data into matrix data through a sequence delay reconstruction module, and then completing downsampling through a convolution combination module Conv by using a convolution kernel with the size of 6x6 and the step length of 2, wherein the length and the width of the matrix are reduced to be 1/2 of the original length and the width of the matrix; the convolution combination module Conv consists of a standard two-dimensional convolution module Conv2d, a normalization module BatchNorm2d and an activation function module SiLU;
step (B2), sequentially passing through a plurality of groups of cross convolution splicing combination modules CSPcat and Conv modules, wherein the specific passing steps are as follows,
step (B21), when the data passes through the splicing and combining module CSPcat, two groups of data are obtained by two groups of modules Conv with the size of 1x1, the step length of 1 convolution kernel and the output channel number halved, wherein one group of data sequentially passes through the modules Conv with the size of 1x1, the step length of 1 and the output channel number halved, the modules Conv with the step length of 1 and the output channel number doubled, and the other group of data does not do any processing and do splicing operation;
step (B22), the data continue to pass through a normalization module BatchNorm2d, an activation function module SiLU and a module Conv with the size of 1x1 and the step length of 1 to obtain final output data, wherein the dimension of the output data is the same as that of the input data, an activation function formula used by the activation function module SiLU is shown as a formula (3),
f(x)＝x*sigmoid(x)＝x/(1+e -x )(3)。
the data is subjected to data distribution change in a module Conv through a normalization module BatchNorm2d module so as to accelerate the convergence speed and stability of a network, and the module does not change the dimension of the data; compared with the traditional Relu function, the SiLU function can improve the estimation accuracy of the network through the activation function module, and the module does not change the dimension of data;
training by using the reconstructed matrix M data and the CSP convolutional neural network to obtain optimal network parameters, wherein the training comprises the specific steps of converting the matrix M data into a CSP convolutional neural network internal convolutional combination module Conv, a splicing combination module CSPcat, a flattening layer flat, a Linear layer Linear and a dimensional compression function sequence.
The data and the data set for training and prediction mainly comprise data with decomposition voltage or current at the power inlet and data of the voltage or current of each electric appliance; the data are time series data with time information.
And (D) arranging a network at the monitoring end, loading corresponding parameters, and decomposing input data to obtain the working state of the load electric appliance, wherein the data dimension required by specific application is adapted by modifying the parameters of the last layers of the network.
To further illustrate the invention, a specific embodiment of the invention is described below, including the following steps,
in the first step, assuming that l=64 is taken, the time sequence length of the input network is 2xl=128, and after the data is input into the network, as shown in fig. 1, the data needs to be converted into a matrix M by a transmat module, where the data dimension is 2×64×64;
second, 2x64x64 dimension data is passed through a convolution combination module Conv, and parameters of the module are set to be (2,32,6,2,2);
taking Conv (2,32,6,2,2) as an example, a standard two-dimensional convolution module Conv2d finishes convolution and downsampling, wherein 2 is the number of input channels, 32 is the number of output channels, 6 is a 6x6 convolution kernel, the first 2 represents step length 2, the second 2 represents zero filled in 2 units at the periphery of the data matrix, and when the zero filling unit defaults to 1, the parameter is omitted;
fourth, after the data passes through the module Conv (2,32,6,2,2), the dimension is changed from 2x64x64 to 32x32x32, then the data in the dimension of 32x32x32 is passed through the convolution combination module CSPcat, the parameters of the module are set to 32,32,1, the first 32 represents the number of channels of the input data, and the second 32 represents the number of channels of the output data;
fifthly, when data passes through a splicing and combining module CSPcat, firstly, two groups of data with 16x32x32 dimensions are respectively obtained through two groups of modules Conv with the size of 1x1 and the step length of 1 convolution kernel and the output channel number halved; the data of 16x32x32 dimension is converted into data of 8x32x32 dimension by a module Conv with 1x1 step length and halved output channel number, then converted into data of 16x32x32 dimension by a module Conv with 3x3 step length and 1 step length and doubled output channel number, wherein a parameter 1 in a convolution combination module CSPcat (32,32,1) represents that the process is repeated for 1 time;
sixthly, performing splicing operation on the group of 16x32x 32-dimensional data and the other group of 16x32x 32-dimensional data, converting the data in the 32x32x 32-dimensional data into the data in the 32x32x 32-dimensional data, and then continuously passing the data in the 32x32x 32-dimensional data through a normalization module BatchNorm2d, an activation function module SiLU and a module Conv with the size of 1x1 and the step length of 1 to obtain final output data, wherein the channel number of the output data is 32;
seventh, the data is passed through a module Conv (32,64,3,2), the original data size is downsampled to 16x16, the channel number is increased to 64, then the data is passed through a CSPcat (64,64,3) module, and the data dimension is still 64x16x16;
eighth step, the data is processed through a module Conv (64,128,3,2) to enable the original data size to be downsampled to 8x8, and the channel number is increased to 128; the data is passed through a CSPcat (128,128,3) module, and the data dimension is 128x8x8;
ninth, downsampling the data to be 4x4 by a module Conv (128,32,3,2), and reducing the channel number to be 32 to obtain data with the dimension of 32x4x 4;
tenth, flattening the latter two dimensions of the data through the flat (2, 3) to obtain data with the dimension of 32x16, and then passing the data with the dimension of 32x16 through a Linear layer Linear (4 x4, 1) to obtain data with the dimension of 32x 1;
eleventh step, converting the data into time sequence data with length of 32 through sequence operation, namely the data of a certain electric appliance obtained after the original data is decomposed, and adapting the data dimension required by specific application through modifying the parameters of the last layers of the network; if the second parameter 32 of the last module Conv (128,32,3,2) is modified to 128, the length of the final output data can be modified to 128.
Experiments based on a uk-dale dataset prove that the comprehensive performance of the algorithm (ReDeSeq) is superior to that of classical algorithms such as DAE, seq2Point, biLSTM and SGN in terms of estimation precision, recall rate, f1 fraction and the like, and the comprehensive performance is shown in tables 1-3.
Table 1 estimation accuracy
| Precision | DAE | Seq2Point | BiLSTM | SGN | ReDeSeq |
| Fridge | 0.809182 | 0.844878 | 0.829618 | 0.923519 | 0.857839 |
| Kettle | 0.99726 | 0.984 | 0.997573 | 0.995984 | 0.995157 |
| Washing machine | 0.62172 | 0.751305 | 0.516338 | 0.725113 | 0.76441 |
Table 2 recall rate
| Recall | DAE | Seq2Point | BiLSTM | SGN | ReDeSeq |
| Fridge | 0.657365 | 0.627075 | 0.664549 | 0.790657 | 0.751576 |
| Kettle | 0.742857 | 0.753061 | 0.838776 | 0.506122 | 0.838776 |
| Washing machine | 0.942571 | 0.986038 | 0.965753 | 0.972866 | 0.960748 |
TABLE 3f1 score
| f1score | DAE | Seq2Point | BiLSTM | SGN | ReDeSeq |
| Fridge | 0.725416 | 0.719863 | 0.737965 | 0.851939 | 0.801199 |
| Kettle | 0.851462 | 0.853179 | 0.911308 | 0.671177 | 0.910299 |
| Washing machine | 0.749241 | 0.852814 | 0.672907 | 0.830915 | 0.851407 |
In summary, according to the non-invasive load monitoring method based on the sequence delay reconstruction CSP convolutional neural network, firstly, the 2xL length time sequence in the data set is reconstructed into the LxLx2 dimension matrix M through sequence delay by utilizing the characteristic of array signal processing, then the convolutional neural network is constructed in a targeted manner by utilizing the CSP network characteristic according to the characteristics of the matrix M, then the reconstructed matrix M data and the CSP convolutional neural network are used for training and obtaining optimal network parameters, then a network is arranged at a monitoring end, corresponding parameters are loaded, and then the input data is decomposed, so that the working state of a load electrical appliance is obtained; the method has the advantages that the influence of time sequence change in data on a monitoring result is enhanced, meanwhile, the methods such as feature map splitting and residual network connection are utilized, the network accuracy is improved while the parameter quantity of a network is reduced, the time sequence relation and state change in the data are considered during monitoring, the monitoring efficiency is guaranteed while the monitoring precision is improved, a user or enterprise electricity utilization data set is obtained at an electric power inlet through low-cost electric power monitoring equipment such as an intelligent ammeter and the like, the designed network is trained by using the data set, and the type and state of the actually used electricity load of the user or enterprise can be obtained only by obtaining the low-frequency electricity utilization data of the user or enterprise, so that the reasonable and safe electricity utilization of the user or enterprise is guided in a targeted manner, and an electric power supply department is assisted to perfect electric power dispatching work.
The foregoing has outlined and described the basic principles, features, and advantages of the present invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, and that the above embodiments and descriptions are merely illustrative of the principles of the present invention, and various changes and modifications may be made without departing from the spirit and scope of the invention, which is defined in the appended claims. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (4)
1. The non-invasive load monitoring method based on the sequence delay reconstruction CSP convolutional neural network is characterized by comprising the following steps of: comprises the steps of,
utilizing the characteristic of array signal processing, reconstructing a 2xL length time sequence in a data set into a matrix M with LxLx2 dimension through sequence delay;
aiming at the characteristics of the matrix M, the specific steps of constructing the convolutional neural network by utilizing the CSP network characteristics are as follows,
step (B1), after data is input into the network, firstly converting the data into matrix data through a sequence delay reconstruction module, and then completing downsampling through a convolution combination module Conv by using a convolution kernel with the size of 6x6 and the step length of 2, wherein the length and the width of the matrix are reduced to be 1/2 of the original length and the width of the matrix; the convolution combination module Conv consists of a standard two-dimensional convolution module Conv2d, a normalization module BatchNorm2d and an activation function module SiLU;
step (B2), sequentially passing through a plurality of groups of cross convolution splicing combination modules CSPcat and Conv modules, wherein the specific passing steps are as follows,
step (B21), when the data passes through the splicing and combining module CSPcat, two groups of data are obtained by two groups of modules Conv with the size of 1x1, the step length of 1 convolution kernel and the output channel number halved, wherein one group of data sequentially passes through the modules Conv with the size of 1x1, the step length of 1 and the output channel number halved, the modules Conv with the step length of 1 and the output channel number doubled, and the other group of data does not do any processing and do splicing operation;
step (B22), the data continue to pass through a normalization module BatchNorm2d, an activation function module SiLU and a module Conv with the size of 1x1 and the step length of 1 to obtain final output data, wherein the dimension of the output data is the same as that of the input data, an activation function formula used by the activation function module SiLU is shown as a formula (3),
f(x)＝x*sigmoid(x)＝x/(1+e -x ) (3)；
training by using the reconstructed matrix M data and the CSP convolutional neural network to obtain optimal network parameters;
and (D) arranging a network at the monitoring end, loading corresponding parameters, and decomposing input data to obtain the working state of the load electric appliance.
2. The non-invasive load monitoring method based on sequence delay reconstruction CSP convolutional neural network of claim 1, wherein: a step (A) of reconstructing a 2xL length time sequence in a data set into a matrix M in LxLx2 dimension through sequence delay by utilizing the characteristic of array signal processing, which comprises the following specific steps,
step (A1), a sequence with the length of 2xL is cut off into a sequence with the length of L by a time unit of sequence delay, and the total L+1 sequences are obtained;
reconstructing the first L sequences into an LxL dimension matrix A, reconstructing the last L sequences into an LxL dimension matrix B, subtracting the matrix A from the matrix B to obtain an LxL dimension matrix C, and splicing the matrix C and the matrix A into A2 xLxL matrix M; wherein the specific steps of the construction of the matrix M are as follows,
step (A22), the sequence of 2xL length is set as [ a ] 1 a 2 …a 2*L ]Then the LxL dimension matrix A is obtained through transformationAnd B and C, and as shown in formula (1),
wherein a is i Representing elements in the sequence, the subscripts of which represent the position of the data in the original sequence, i.e. a i A being the first element of the ith sequence, a i+L-1 Is the last element in the original sequence;
step (A23), the matrix M is A2 xLxL-dimensional matrix formed by splicing an LxL-dimensional matrix C and an LxL-dimensional matrix A, as shown in a formula (2),
3. the non-invasive load monitoring method based on sequence delay reconstruction CSP convolutional neural network of claim 2, wherein: training by using the reconstructed matrix M data and the CSP convolutional neural network to obtain optimal network parameters, wherein the training comprises the specific steps of converting the matrix M data into a CSP convolutional neural network internal convolutional combination module Conv, a splicing combination module CSPcat, a flattening layer flat, a Linear layer Linear and a dimensional compression function sequence.
4. A method for non-invasive load monitoring based on sequence delay reconstruction CSP convolutional neural network of claim 3, wherein: and (D) arranging a network at the monitoring end, loading corresponding parameters, and decomposing input data to obtain the working state of the load electric appliance, wherein the data dimension required by specific application is adapted by modifying the parameters of the last layers of the network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210259630.5A CN114745299B (en) | 2022-03-16 | 2022-03-16 | Non-invasive load monitoring method based on sequence delay reconstruction CSP convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210259630.5A CN114745299B (en) | 2022-03-16 | 2022-03-16 | Non-invasive load monitoring method based on sequence delay reconstruction CSP convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114745299A CN114745299A (en) | 2022-07-12 |
| CN114745299B true CN114745299B (en) | 2023-06-13 |
Family
ID=82277602
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210259630.5A Active CN114745299B (en) | 2022-03-16 | 2022-03-16 | Non-invasive load monitoring method based on sequence delay reconstruction CSP convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114745299B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116341616B (en) * | 2022-11-11 | 2023-10-17 | 南京工程学院 | Electric load information acquisition method based on matrix reconstruction two-dimensional convolution network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| SU1352662A1 (en) * | 1985-04-30 | 1987-11-15 | Военный Инженерный Краснознаменный Институт Им.А.Ф.Можайского | Device for retrieval by delay of combination pseudorandom sequences |
| CN111012336A (en) * | 2019-12-06 | 2020-04-17 | 重庆邮电大学 | Parallel convolutional network motor imagery electroencephalogram classification method based on spatio-temporal feature fusion |
| CN113902102A (en) * | 2021-10-19 | 2022-01-07 | 南京工程学院 | Non-invasive load decomposition method based on seq2seq |
| CN113970667A (en) * | 2021-10-10 | 2022-01-25 | 上海梦象智能科技有限公司 | Non-invasive load monitoring method based on midpoint value of prediction window |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20220038902A1 (en) * | 2020-11-13 | 2022-02-03 | Markus Dominik Mueck | Technologies for radio equipment cybersecurity and multiradio interface testing |
-
2022
- 2022-03-16 CN CN202210259630.5A patent/CN114745299B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| SU1352662A1 (en) * | 1985-04-30 | 1987-11-15 | Военный Инженерный Краснознаменный Институт Им.А.Ф.Можайского | Device for retrieval by delay of combination pseudorandom sequences |
| CN111012336A (en) * | 2019-12-06 | 2020-04-17 | 重庆邮电大学 | Parallel convolutional network motor imagery electroencephalogram classification method based on spatio-temporal feature fusion |
| CN113970667A (en) * | 2021-10-10 | 2022-01-25 | 上海梦象智能科技有限公司 | Non-invasive load monitoring method based on midpoint value of prediction window |
| CN113902102A (en) * | 2021-10-19 | 2022-01-07 | 南京工程学院 | Non-invasive load decomposition method based on seq2seq |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114745299A (en) | 2022-07-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109543830B (en) | Splitting accumulator for convolutional neural network accelerator | |
| CN109635936A (en) | A kind of neural networks pruning quantization method based on retraining | |
| CN114745299B (en) | Non-invasive load monitoring method based on sequence delay reconstruction CSP convolutional neural network | |
| CN106162868A (en) | High efficiency indoor localization method based on location fingerprint | |
| CN111652360B (en) | Convolution operation device based on pulsation array | |
| CN112051480A (en) | Neural network power distribution network fault diagnosis method and system based on variational modal decomposition | |
| Chitsaz et al. | Acceleration of convolutional neural network using fft-based split convolutions | |
| CN113159226B (en) | Inverter fault diagnosis method with integration of depth features and statistical features | |
| CN116304889A (en) | Receptor classification method based on convolution and transducer | |
| Moon et al. | FPGA-based sparsity-aware CNN accelerator for noise-resilient edge-level image recognition | |
| CN113807497B (en) | Unpaired image translation method for enhancing texture details | |
| CN107957977A (en) | A kind of computational methods and Related product | |
| Zhang et al. | Short-term wind speed forecasting based on adaptive secondary decomposition and robust temporal convolutional network | |
| CN116341616B (en) | Electric load information acquisition method based on matrix reconstruction two-dimensional convolution network | |
| CN110196735A (en) | A kind of computing device and Related product | |
| CN106066941B (en) | A kind of electromagnetic scattering rapid analysis method based on CBFM and SMW algorithms | |
| CN109495113A (en) | A kind of compression method and device of EEG signals | |
| Fan et al. | Hardware and Algorithm Co-Optimization for pointwise convolution and channel shuffle in ShuffleNet V2 | |
| CN113158134B (en) | Method, device and storage medium for constructing non-invasive load identification model | |
| CN111476408B (en) | Power communication equipment state prediction method and system | |
| CN117114712A (en) | Power utilization side carbon monitoring method, model training method, device and computer equipment | |
| CN110196734A (en) | A kind of computing device and Related product | |
| Chen et al. | H-RIS: Hybrid Computing-in-Memory Architecture Exploring Repetitive Input Sharing | |
| CN111126593B (en) | Reconfigurable natural language deep convolutional neural network accelerator | |
| CN110390391B (en) | Mapping device and method based on three-dimensional convolutional neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |