CN117807528A - Non-invasive household appliance state identification method and medium based on long-short-term memory network - Google Patents
Non-invasive household appliance state identification method and medium based on long-short-term memory network Download PDFInfo
- Publication number
- CN117807528A CN117807528A CN202311864845.0A CN202311864845A CN117807528A CN 117807528 A CN117807528 A CN 117807528A CN 202311864845 A CN202311864845 A CN 202311864845A CN 117807528 A CN117807528 A CN 117807528A
- Authority
- CN
- China
- Prior art keywords
- model
- long
- data set
- training
- lstm
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 52
- 230000015654 memory Effects 0.000 title claims abstract description 27
- 238000012549 training Methods 0.000 claims abstract description 61
- 230000008447 perception Effects 0.000 claims abstract description 35
- 238000013140 knowledge distillation Methods 0.000 claims abstract description 15
- 238000013139 quantization Methods 0.000 claims abstract description 11
- 238000013527 convolutional neural network Methods 0.000 claims description 24
- 230000006870 function Effects 0.000 claims description 13
- 230000004913 activation Effects 0.000 claims description 6
- 238000007667 floating Methods 0.000 claims description 6
- 238000005516 engineering process Methods 0.000 claims description 5
- 238000004458 analytical method Methods 0.000 claims description 4
- 238000004590 computer program Methods 0.000 claims description 3
- 230000000737 periodic effect Effects 0.000 claims description 3
- 238000005265 energy consumption Methods 0.000 abstract description 2
- 230000006835 compression Effects 0.000 description 9
- 238000007906 compression Methods 0.000 description 9
- 230000005611 electricity Effects 0.000 description 9
- 238000012545 processing Methods 0.000 description 8
- 238000012360 testing method Methods 0.000 description 7
- 230000006399 behavior Effects 0.000 description 6
- 230000008569 process Effects 0.000 description 6
- 230000008859 change Effects 0.000 description 5
- 230000007547 defect Effects 0.000 description 4
- 230000007774 longterm Effects 0.000 description 4
- 238000012544 monitoring process Methods 0.000 description 4
- 238000011176 pooling Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000009434 installation Methods 0.000 description 3
- 238000012423 maintenance Methods 0.000 description 3
- 230000003287 optical effect Effects 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 230000002123 temporal effect Effects 0.000 description 3
- 230000001052 transient effect Effects 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000007405 data analysis Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 230000000644 propagated effect Effects 0.000 description 2
- 238000011002 quantification Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000032683 aging Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 239000000835 fiber Substances 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000036962 time dependent Effects 0.000 description 1
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a non-invasive household appliance state identification method based on a long-short-term memory network, which comprises the following steps: collecting a single-electric-appliance current value and constructing a training data set; training a perception model based on CNN-LSTM by using a training data set, wherein the perception model adopts a CNN module as a feature extractor, adopts the LSTM module to separate the electrical signals of each electrical appliance from the total signals, and adopts a decision layer to determine the probability of the switching state of the electrical appliance according to the separated electrical signals; compressing the trained perception model by adopting a knowledge distillation and model quantization method to obtain a lightweight model; and deploying the lightweight model to a mobile terminal, and identifying the state of the electric appliance in real time. The invention has the advantages of low energy consumption, low cost, simple configuration, high recognition precision, high reasoning speed and the like.
Description
Technical Field
The invention belongs to the field of current identification, and particularly relates to a non-invasive household appliance state identification method and medium based on a long-short time memory network.
Background
In the household electrical appliance load identification method, the traditional invasive method mainly has the following defects: (1) The traditional invasive method has the defects of high cost, complex installation and maintenance and the like because the acquisition sensing device is required to be installed in the household appliance, and the method only needs to acquire the total information of the electricity consumption of the user at the electric power inlet and then performs data analysis and prediction on the total electricity consumption data, and has the advantages of low cost, convenience in installation and maintenance and the like. (2) The current household appliance state identification algorithm mostly adopts a time domain method and a time frequency method, the time domain method and the time frequency method mostly select high-frequency signals of steady state or transient state electric parameters of the household appliance, such as steady state voltage, current, power parameters and the like, as load labels, and the household appliance state identification algorithm has the defects of large acquired data volume, large occupied storage space and the like. (3) The traditional household appliance state recognition algorithm mostly adopts a k-nearest neighbor algorithm, an LMD (local mean decomposition), an HHT (Hilbert transform), a wavelet transform and other traditional clustering and decomposing algorithms to realize the household appliance switch state judgment, and has the defects of higher operation space and time complexity, non-ideal recognition accuracy under the environments of multiple household appliances, complex multistable and transient signals and the like.
With the development of artificial intelligence and intelligent home, home appliance state identification research has become an important research topic in the fields of electricity consumption behavior analysis, household electricity consumption safety, reasonable power distribution electricity consumption and the like. Based on the existing load identification research, a user can monitor the running state of the electric appliance and autonomously optimize the electricity consumption. Further, the electricity utilization behavior of the user can be analyzed and predicted according to different household appliance state data. However, too many sensors may be involved in infringement of user privacy, complicated installation and maintenance, high cost, and the like. The traditional non-supervision learning algorithm is not high in state accuracy of identifying single household appliances from electrical signals of multi-household appliance mixed and complex electrical steady state and transient state information, a large number of data samples are needed for training during supervision learning, the number of required training sets is exponentially increased along with the increase of the types of the electrical appliances, and the state identification by utilizing the traditional non-supervision learning algorithm becomes impractical.
Disclosure of Invention
The invention provides a non-invasive household appliance state identification method based on a long-short-term memory network.
The technical scheme for realizing the purpose of the invention is as follows: a non-invasive household appliance state identification method based on a long-short-term memory network comprises the following specific steps:
collecting a single-electric-appliance current value and constructing a training data set;
training a perception model based on CNN-LSTM by using a training data set, wherein the perception model adopts a CNN module as a feature extractor, adopts the LSTM module to separate the electrical signals of each electrical appliance from the total signals, and adopts a full connection layer to determine the probability of the switching state of the electrical appliance according to the separated electrical signals;
compressing the trained perception model by adopting a knowledge distillation and model quantization method to obtain a lightweight model;
and deploying the lightweight model to a mobile terminal, and identifying the state of the electric appliance in real time.
Preferably, the specific method for collecting the current value of the single electric appliance and constructing the training data set is as follows:
collecting single electric appliancesCurrent value, according to kirchhoff's law, single electric appliance current is combined, n kinds of electric appliances are combined into 2 n As a training data set.
Preferably, gaussian-distributed random noise is added to the training data set according to the mean and variance of the training data set, and gaussian noise is generated and added to the current signal by setting the mean and standard deviation.
Preferably, the CNN module includes three layers of CNNs, a ReLU activation function and a pooled layer enhancement network, each layer of CNNs includes a series of convolution operations for extracting time patterns, frequency information and amplitude variation characteristics in the current data, and feature dimensions are reduced by nonlinear characteristics of the ReLU activation function and the pooled layer enhancement network.
Preferably, the LSTM module is configured to perform a time-series evolution analysis on the features extracted by the CNN module, to obtain a significant time dynamics in the current signal and a critical event, where the significant time dynamics includes a periodic fluctuation, a spike, or a drop of the current signal, and the critical event includes turning on or off of the electrical appliance.
Preferably, a binary cross entropy loss function is employed as the loss function for CNN-LSTM based perceptual model training.
Preferably, the training perception model is compressed by adopting a knowledge distillation and model quantization method, and the specific method for obtaining the lightweight model is as follows:
the trained perception model based on CNN-LSTM is used as a teacher model, key information of the teacher model is transmitted to a student model by adopting a knowledge distillation technology, the student model is trained, and the number of CNN layers and LSTM layers of the student model is less than that of the teacher model;
and converting the weight of the student model into an integer, and taking the trained student model with the weight being the integer as a lightweight model.
Preferably, the specific method for transmitting the key information of the teacher model to the student model and training the student model by adopting the knowledge distillation technology is as follows:
training the training data set of the perception model, the corresponding label and the perception model of the perception model to train the soft target output and input student model with the probability distribution form of the same training data set until the model converges.
Compared with the prior art, the invention has the remarkable advantages that: the invention provides a non-invasive household appliance state recognition algorithm based on a CNN-LSTM model, which has the advantages of low energy consumption, low cost, simple configuration, high recognition precision, high reasoning speed and the like;
the invention provides a method for constructing a data set by utilizing single household appliance current, which is used for generating the data set by arranging and combining all single household appliance current values based on kirchhoff's law and generating a new data set through online learning based on user feedback.
According to the method, the problem of algorithm operation complexity is considered, and a model compression method combining knowledge distillation and model quantization is adopted, so that the reasoning speed of the model is greatly improved under the condition that the algorithm identification accuracy is not obviously reduced.
The present invention will be described in further detail with reference to the accompanying drawings.
Drawings
Fig. 1 is a flowchart of a non-invasive home appliance state recognition method based on a long-short-term memory network according to the present invention.
FIG. 2 is a schematic diagram of a perception model based on CNN-LSTM.
Detailed Description
A non-invasive household appliance state identification method based on a long-short-term memory network comprises the following specific steps:
collecting a single-electric-appliance current value and constructing a training data set;
taking the characteristics of characteristic signals of the electrical appliances into consideration, training a perception model based on CNN-LSTM by using a training data set, training a model for each electrical appliance independently, and training n models for n electrical appliances, wherein the perception model adopts a CNN network as a characteristic extractor, and adopts an LSTM network to separate the electrical signals of each electrical appliance from the total signals;
compressing the trained perception model by adopting a knowledge distillation and model quantization method to obtain a lightweight model;
and deploying the lightweight model to a mobile terminal, and identifying the state of the electric appliance in real time.
As an example, the current values of the single electric appliances are collected, the single electric appliance currents are combined according to kirchhoff's law, and 2 n kinds of mixed currents are combined as training data sets in the case of n kinds of electric appliances. The supervision and learning need a large amount of data to train, in the field of home state identification, a training data set needs to cover the switch state combination condition of all electric appliances under the condition of multiple electric appliances, so that the training data acquisition is too complicated, a newly added electric appliance needs to acquire the data set again, and the training time is too long due to the fact that the data set is too large. The training set construction method of the embodiment solves the problems.
Specifically, a multi-household appliance scene is constructed, and the current is collected and monitored by combining a singlechip with a current transformer. The current transformer is responsible for detecting the current and converting the current into an analog signal which can be processed by the singlechip. Each current signal consists of 320 acquisition points, and is converted into a digital signal through an analog-to-digital converter (ADC) on the singlechip. The singlechip is responsible for reading and analyzing the data points, and further transmitting and displaying the data through the corresponding interfaces. The configuration enables the system to accurately monitor the current change, and is suitable for various current monitoring application scenes. After the current signals of each electric appliance are collected, according to kirchhoff's law, the total current value of the parallel electric appliances is equal to the sum of the currents of each electric appliance, and the current signals are arranged and combined to obtain a representative total data set.
Further, adding noise to the training data set generates a comprehensive, updatable data set.
In particular, a multi-appliance scenario is established, in addition to accurate monitoring of current signals, noise data is introduced to simulate current fluctuations in the real world to enhance the diversity and practicality of the data set. Random noise of Gaussian distribution is added according to the mean value and variance of the training data set, gaussian noise is generated by setting proper mean value and standard deviation and added to a current signal, and the noise can simulate real factors such as voltage fluctuation, temporary load change or equipment aging and the like, so that the data set is more comprehensive.
In addition, as time passes and the appliance usage patterns change, the data set is updated through an online learning mechanism, and is continuously adjusted and optimized according to the latest current monitoring data so as to adapt to the new usage patterns and environmental conditions. By periodically re-analyzing the current data and incorporating new observations into the existing data set. The online learning mechanism ensures that the data set is always kept up to date, so that the prediction accuracy and reliability of the monitoring system on the behavior of the electrical appliance are improved.
Further, the training data set adopts a one-hot type two-class label, and if [1,0] represents that the electric appliance is closed, and [0,1] represents that the electric appliance is opened.
The data set generated by the embodiment has the advantages of small data volume, comprehensive combination of the switch states of the covered electric appliances and capability of updating in real time according to user feedback.
As an embodiment, as shown in fig. 2, the perception model includes a cascaded three-layer Convolutional Neural Network (CNN), two-layer long-short-term memory network (LSTM), and a decision layer, where the three-layer convolutional neural network forms a CNN module, and the two-layer long-short-term memory network forms an LSTM module.
Specifically, after each layer of convolutional neural network is connected with a pooling layer, and the last layer of pooling layer outputs time mode, frequency information and amplitude change characteristics in current data, python and tensor low built-in functions are called to perform latitude conversion and tensor segmentation on the characteristics, so that the output of the pooling layer meets the requirement of long-term and short-term memory network input.
Specifically, the CNN module is configured to extract features from input data. In the CNN module, each layer of convolutional neural network consists of a series of convolutional operations, and local features such as time modes, frequency information, amplitude change and the like in the current data are effectively extracted through the convolutional operations. And then reducing feature dimensions through a ReLU activation function and the non-linear characteristic of the pooling layer enhanced network, and finally sending the extracted multidimensional feature map into an LSTM layer.
Specifically, the LSTM module is used for processing time series data and capturing long-term dependency relations therein. LSTM is a special Recurrent Neural Network (RNN) capable of retaining long-term dependency information in sequence data processing. By processing the LSTM layer, the model can better understand the temporal dynamics in the current data and effectively classify or predict the sequence.
Specifically, the primary role of the LSTM module is to capture and understand the long-term time dependence and dynamic changes in current data. By combining the local features (such as temporal patterns, frequency information, and amplitude variations) extracted by the CNN module, the LSTM module can deeply analyze the evolution of these features throughout the time sequence, capturing sharply important temporal dynamics in the current signal, such as periodic fluctuations, spikes, or dips, as well as critical events, such as the turning on and off of the appliance. After processing the current data, the long-short term memory network generates a comprehensive characteristic representation that includes not only the key time-dependent and dynamic characteristics of the current signal, but also the context information within the entire observation window. This enables the output of the LSTM module to provide rich information for further analysis of the current data, which information may be used in the perception model to identify the switching state of the appliance.
Specifically, the decision layer determines the probability of the appliance switching state from the output of the LSTM module, thereby enabling the model to perform accurate data interpretation and prediction based on a comprehensive understanding of the current time series.
Specifically, the decision layer adopts a full connection layer.
Further, the CNN-LSTM-based perception model is trained by adopting binary cross entropy as a loss function of the binary classification problem. The model perception model training uses a back propagation algorithm in combination with a gradient descent method to optimize network parameters. In each iteration, the gradient of the loss function with respect to the model parameters is calculated and the parameters are updated to minimize the loss function.
Furthermore, an Adam optimizer is adopted during training, so that the training accuracy is ensured, and meanwhile, the training process is accelerated and the convergence rate of the model is improved. Training strategies such as learning rate scheduling, early stopping and Dropout are adopted to improve the generalization capability of the model and prevent overfitting.
The invention fully utilizes the advantages of CNN in characteristic extraction and the capability of LSTM in processing long-term dependency relationship in time sequence data based on the perception model of CNN-LSTM, so that the model is more effective and accurate in processing complex current data.
As an embodiment, the trained perception model is compressed by adopting a knowledge distillation and model quantization method to obtain a lightweight model, the model can be reduced to 1/64 of the original model, and the prediction accuracy is not obviously reduced, and the specific method comprises the following steps:
the trained perception model is used as a teacher model through a knowledge distillation technology, and key information of the training model is transmitted to a smaller and lighter student model.
The student model adopts two layers of convolution and one layer of LSTM as the student model, and the parameter quantity is about 1/4 of that of a teacher model, so that smaller and lighter weight can be realized.
The key information includes the soft target output (i.e., probability distribution of the entire output layer) of the teacher model, the feature representation of the middle layer, and the intrinsic rules and abstract representation of the current data.
In order to enable the student model to simulate the behavior and performance of the teacher model while keeping a small volume, the output of the teacher model participates in the training process of the student model, and the training process of the student model specifically comprises the following steps:
and training the training data set of the perception model, the corresponding label and the perception model of the perception model to output and input the soft target with the same probability distribution form of the training data set into the student model. The soft target output includes a confidence level of the teacher model for each class;
the training target of the student model is not only to accurately predict the original label, but also to approximate the soft target of the teacher model as much as possible. In this way, the student model can learn the behavior and decision process of the teacher model, thereby mimicking the advanced behavior and performance of the teacher model while maintaining a small volume. The training process allows the student model to learn simple classification tasks, and can capture deep understanding and fine classification boundaries of the teacher model on data, which is very critical for improving the accuracy and generalization capability of the student model.
Model quantification is then performed on the student model, a technique that reduces the storage space required for the model. The 32-bit floating point number parameter in the student model is converted into the 8-bit low-precision integer representation, so that the storage requirement of the model is remarkably reduced, and the reasoning speed of the model can be improved because the integer operation is more efficient in hardware than the floating point operation. Through the two steps, the student model is reduced to 1/64 of the original model, the prediction accuracy is not obviously reduced, and the reasoning speed is improved by about 4 times.
Further, in a deep learning model in which the trained student model is unquantized, the weights (i.e., parameters of each layer in the neural network) and the activation values (output of the neural network layer) are typically stored in a 32-bit floating point number format. This format provides sufficient accuracy to capture subtle changes in the training process, but relatively occupies more memory and computing resources. The invention adopts a symmetrical quantization strategy, calculates a scaling factor and a zero point to map floating points to an integer range, applies integer range parameters to complete quantization conversion, converts 32-bit floating point number parameters in a student model into 8-bit low-precision integer representation, and takes the student model with converted weight as a lightweight model.
In a further embodiment, a lightweight model is deployed to the smart meter. Electricity meters, which are a typical type of edge computing device, generally have low processing power and limited memory. Model compression is therefore critical to ensure that the model can run smoothly on these devices. The smart electric meter can be provided with a plurality of lightweight models of different electric appliances, and each lightweight model predicts the switching state of the corresponding electric appliance respectively.
Through the series of model optimization and compression strategies, the effective operation of the intelligent electric meter in a resource-limited environment can be realized while the performance of the model is ensured, and powerful data analysis and processing support is provided for applications such as intelligent electric meters.
Taking a model notebook computer in a lightweight model corresponding to 15 electric appliances as an example, the test set consists of 2048 mixed current data, the original model size is 1731k, the size after two model compression is 25.7k, and the test precision on the test set is not obviously reduced, as shown in the table 2:
| Method | model size (front/rear) | Test precision (front/back) |
| Knowledge distillation | 1731k/130k | 99.7％/94.92％ |
| Quantization | 1731k/185k | 99.7％/95.05％ |
| Knowledge distillation + quantification | 1731k/25.7k | 99.7％/94.7％ |
TABLE 1
In order to compare the reasoning speed of the model before and after compression, the model before and after compression is tested on a test set and the test time is calculated, so as to obtain the time consumption pairs before and after model compression as shown in table 3:
| index (I) | Before model compression | After the model is compressed |
| Time consuming testing | 16.9s | 4.4s |
TABLE 2
As one embodiment, the device for collecting the electricity consumption information of the user is a smart electric meter at the incoming line of the user, the smart electric meter collects the single-electric-appliance current values of 15 electric appliances, and a training data set is generated through an algorithm. And training a CNN-LSTM model by the upper computer, deploying the trained model on the intelligent ammeter through model compression, predicting the state of the household appliance in real time, and correcting a model prediction result through online learning and updating a training set.
As an embodiment, a computer readable storage medium has a computer program stored thereon, where the program when executed by a processor may implement any of the method for sensing a state of a powered device based on a graph convolutional neural network in the above embodiment.
The computer storage media of embodiments of the invention may take the form of any combination of one or more computer-readable media. The computer readable medium may be a computer readable signal medium or a computer readable storage medium. The computer readable storage medium may be, for example, but not limited to: an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or a combination of any of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium would include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In this document, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.
The computer readable signal medium may include a propagated data signal with computer readable program code embodied therein, either in baseband or as part of a carrier wave. Such a propagated data signal may take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination of the foregoing. A computer readable signal medium may also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.
Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to:
wireless, wire, fiber optic cable, RF, etc., or any suitable combination of the foregoing.
Computer program code for carrying out operations of the present invention may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, smalltalk, C ++ and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any kind of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or may be connected to an external computer (for example, through the Internet using an Internet service provider).
Claims (10)
1. A non-invasive household appliance state identification method based on a long-short-time memory network is characterized by comprising the following specific steps:
collecting a single-electric-appliance current value and constructing a training data set;
training a perception model based on CNN-LSTM by using a training data set, wherein the perception model adopts a CNN module as a feature extractor, adopts the LSTM module to separate the electrical signals of each electrical appliance from the total signals, and adopts a decision layer to determine the probability of the switching state of the electrical appliance according to the separated electrical signals;
compressing the trained perception model by adopting a knowledge distillation and model quantization method to obtain a lightweight model;
and deploying the lightweight model to a mobile terminal, and identifying the state of the electric appliance in real time.
2. The non-invasive home appliance state recognition method based on long and short time memory network according to claim 1, wherein the specific method for collecting the single electric appliance current value and constructing the training data set is as follows:
collecting the current value of a single electric appliance, combining the current of the single electric appliance according to kirchhoff's law, wherein n electric appliances are combined into 2 n As a training data set.
3. The non-invasive home appliance state recognition method based on long and short time memory network according to claim 2, wherein gaussian distributed random noise is added to the training data set according to the mean and variance of the training data set, and gaussian noise is generated by setting the mean and standard deviation and added to the current signal.
4. The non-invasive home appliance state recognition method based on long and short time memory network according to claim 1, wherein the CNN module comprises three layers of CNNs, a ReLU activation function and a pooled layer enhancement network in cascade, each layer of CNNs comprises a series of convolution operations for extracting time patterns, frequency information and amplitude variation characteristics in current data, and feature dimensions are reduced by nonlinear characteristics of the ReLU activation function and the pooled layer enhancement network.
5. The non-invasive home appliance state identification method based on long and short time memory network according to claim 1, wherein the LSTM module is configured to perform time-series evolution analysis on features extracted by the CNN module, to obtain important time dynamics in a current signal and a critical event, where the important time dynamics includes periodic fluctuation, spike or drop of the current signal, and the critical event includes turning on or off of an electrical appliance.
6. The non-invasive home appliance state identification method based on long and short time memory network according to claim 1, wherein a binary cross entropy loss function is adopted as a loss function based on CNN-LSTM perception model training.
7. The non-invasive home appliance state identification method based on long and short time memory network according to claim 1, wherein the specific method for compressing the trained perception model by adopting knowledge distillation and model quantization method to obtain the lightweight model is as follows:
the trained perception model based on CNN-LSTM is used as a teacher model, key information of the teacher model is transmitted to a student model by adopting a knowledge distillation technology, the student model is trained, and the number of CNN layers and LSTM layers of the student model is less than that of the teacher model;
and converting the floating point number parameters of the trained student model into integers, and taking the student model with the converted parameters as a lightweight model.
8. The non-invasive home appliance state recognition method based on long and short time memory network according to claim 1, wherein the specific method for transmitting the key information of the teacher model to the student model and training the student model by adopting the knowledge distillation technology is as follows:
training the training data set of the perception model, the corresponding label and the perception model of the perception model to train the soft target output and input student model with the probability distribution form of the same training data set until the model converges.
9. The non-invasive home appliance state recognition method based on long and short time memory network according to claim 1, wherein the electric appliance state result recognized in real time by the lightweight model is compared with the user feedback, if the electric appliance state result is inconsistent with the user feedback result, the training set is updated by online learning, and the perception model based on CNN-LSTM is retrained.
10. A storage medium having stored thereon a computer program, which when executed by a processor implements a non-invasive home appliance state identification method based on a long and short time memory network according to any of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311864845.0A CN117807528A (en) | 2023-12-29 | 2023-12-29 | Non-invasive household appliance state identification method and medium based on long-short-term memory network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311864845.0A CN117807528A (en) | 2023-12-29 | 2023-12-29 | Non-invasive household appliance state identification method and medium based on long-short-term memory network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117807528A true CN117807528A (en) | 2024-04-02 |
Family
ID=90425230
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311864845.0A Pending CN117807528A (en) | 2023-12-29 | 2023-12-29 | Non-invasive household appliance state identification method and medium based on long-short-term memory network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117807528A (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110121988A1 (en) * | 2009-11-20 | 2011-05-26 | Institute For Information Industry | Method and system for recognizing statuses of electric appliances |
| CN110188826A (en) * | 2019-05-31 | 2019-08-30 | 湖南科技大学 | Household electrical appliance operating status non-invasive inspection methods based on intelligent electric meter data |
| WO2021047286A1 (en) * | 2019-09-12 | 2021-03-18 | 华为技术有限公司 | Text processing model training method, and text processing method and apparatus |
| KR20220161849A (en) * | 2021-05-31 | 2022-12-07 | 주식회사 스마트에버 | Method and Apparatus for Predicting Appliance Power Usage State Using Non Intrusive Load Monitoring Based on Long-Short Term Memory |
| CN115527213A (en) * | 2022-09-20 | 2022-12-27 | 上海弘玑信息技术有限公司 | Document character recognition method and electronic equipment |
-
2023
- 2023-12-29 CN CN202311864845.0A patent/CN117807528A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110121988A1 (en) * | 2009-11-20 | 2011-05-26 | Institute For Information Industry | Method and system for recognizing statuses of electric appliances |
| CN110188826A (en) * | 2019-05-31 | 2019-08-30 | 湖南科技大学 | Household electrical appliance operating status non-invasive inspection methods based on intelligent electric meter data |
| WO2021047286A1 (en) * | 2019-09-12 | 2021-03-18 | 华为技术有限公司 | Text processing model training method, and text processing method and apparatus |
| KR20220161849A (en) * | 2021-05-31 | 2022-12-07 | 주식회사 스마트에버 | Method and Apparatus for Predicting Appliance Power Usage State Using Non Intrusive Load Monitoring Based on Long-Short Term Memory |
| CN115527213A (en) * | 2022-09-20 | 2022-12-27 | 上海弘玑信息技术有限公司 | Document character recognition method and electronic equipment |
Non-Patent Citations (3)
| Title |
|---|
| 何恒靖;王灏;肖勇;张恒;赵云;周东国;: "基于Bi-LSTM算法的非侵入式负荷监测模型", 南方电网技术, no. 02, 20 February 2019 (2019-02-20), pages 26 - 32 * |
| 岳建任 等: "基于seq2seq 模型的非侵入式负荷分解算法", 《电测与仪表》, 15 July 2021 (2021-07-15), pages 1 - 10 * |
| 王琪凯: "基于深度学习模型的非侵入式电力负荷分解方法研究", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》, 16 January 2023 (2023-01-16), pages 11 - 35 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| De Baets et al. | Appliance classification using VI trajectories and convolutional neural networks | |
| Dou et al. | Hybrid model for renewable energy and loads prediction based on data mining and variational mode decomposition | |
| Quek et al. | Load disaggregation using one-directional convolutional stacked long short-term memory recurrent neural network | |
| Mian Qaisar | Signal-piloted processing and machine learning based efficient power quality disturbances recognition | |
| KR20200128232A (en) | Apparatus for predicting electricity demand and method thereof | |
| CN114023354A (en) | Guidance type acoustic event detection model training method based on focusing loss function | |
| CN111160626A (en) | Power load time sequence control method based on decomposition and fusion | |
| Ouyang et al. | Model of selecting prediction window in ramps forecasting | |
| Shen et al. | An EMD-RF based short-term wind power forecasting method | |
| CN115659254A (en) | Power quality disturbance analysis method for power distribution network with bimodal feature fusion | |
| Liu et al. | Non-intrusive energy estimation using random forest based multi-label classification and integer linear programming | |
| US20140278338A1 (en) | Stream input reduction through capture and simulation | |
| Qu et al. | A parameter and resolution adaptive algorithm for rapid detection of ramp events in different timescale databases of the power system | |
| CN110866840A (en) | Database modeling method for power load characteristic quantity training based on knowledge graph | |
| Ouyang et al. | Optimisation of time window size for wind power ramps prediction | |
| Fu et al. | Predicting peak day and peak hour of electricity demand with ensemble machine learning | |
| Du et al. | A hierarchical power system transient stability assessment method considering sample imbalance | |
| CN116979503A (en) | Commercial building load prediction method and system based on Informir network | |
| CN116561569A (en) | Industrial power load identification method based on EO feature selection and AdaBoost algorithm | |
| CN117807528A (en) | Non-invasive household appliance state identification method and medium based on long-short-term memory network | |
| CN111090679A (en) | Time sequence data representation learning method based on time sequence influence and graph embedding | |
| Chen et al. | A dual attention-based CNN-GRU model for short-term electric load forecasting | |
| CN115034586A (en) | Method and device for predicting power failure risk of overhead line in strong convection weather | |
| Zheng et al. | Offshore wind power ramp prediction based on optimal combination model | |
| Zhang et al. | Named Entity Recognition for Smart Grid Operation and Inspection Domain using Attention Mechanism |
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 |