CN117473368A - Non-invasive load monitoring semi-supervised learning method and system - Google Patents
Non-invasive load monitoring semi-supervised learning method and system Download PDFInfo
- Publication number
- CN117473368A CN117473368A CN202311426481.8A CN202311426481A CN117473368A CN 117473368 A CN117473368 A CN 117473368A CN 202311426481 A CN202311426481 A CN 202311426481A CN 117473368 A CN117473368 A CN 117473368A
- Authority
- CN
- China
- Prior art keywords
- semi
- supervised learning
- load
- data
- training
- 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- Withdrawn
Links
- 238000000034 method Methods 0.000 title claims abstract description 69
- 238000012544 monitoring process Methods 0.000 title claims abstract description 50
- 230000007246 mechanism Effects 0.000 claims abstract description 27
- 238000012549 training Methods 0.000 claims description 58
- 238000013145 classification model Methods 0.000 claims description 30
- 238000005070 sampling Methods 0.000 claims description 29
- 238000000605 extraction Methods 0.000 claims description 22
- 125000004122 cyclic group Chemical group 0.000 claims description 20
- 230000008569 process Effects 0.000 claims description 20
- 238000012545 processing Methods 0.000 claims description 14
- 230000006870 function Effects 0.000 claims description 10
- 238000012795 verification Methods 0.000 claims description 10
- 238000004590 computer program Methods 0.000 claims description 9
- 238000012216 screening Methods 0.000 claims description 8
- 239000000284 extract Substances 0.000 claims description 7
- 238000002372 labelling Methods 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 5
- 238000001914 filtration Methods 0.000 claims description 4
- 230000001360 synchronised effect Effects 0.000 claims description 4
- 238000012804 iterative process Methods 0.000 claims description 3
- 238000002156 mixing Methods 0.000 claims description 3
- 238000013480 data collection Methods 0.000 abstract description 5
- 238000005096 rolling process Methods 0.000 abstract 1
- 238000004519 manufacturing process Methods 0.000 description 15
- 238000004891 communication Methods 0.000 description 6
- 230000005611 electricity Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 3
- 238000007726 management method Methods 0.000 description 3
- 230000003287 optical effect Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000003908 quality control method Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000004134 energy conservation Methods 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000000835 fiber Substances 0.000 description 1
- 230000003862 health status Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/259—Fusion by voting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a semi-supervised learning method and a semi-supervised learning system for non-invasive load monitoring. The second part of the method is used for an reasoning stage, a consistent voting mechanism is utilized in an optimal interval of a predicted result to obtain a high-precision pseudo tag required by semi-supervised learning, and then the recognition precision is gradually improved through a rolling learning method by a small amount of tagged data and a large amount of untagged data. The non-invasive load monitoring semi-supervised learning method provided by the invention can effectively combine the label-free data to improve the accuracy of the model, and can improve the accuracy and the reliability of the model in the actual scene with limited data collection.
Description
Technical Field
The invention relates to the technical field of deep learning, in particular to a non-invasive semi-supervised learning method and system for load monitoring.
Background
According to the statistics of the International Energy Agency (IEA), the electricity consumption of the industrial industry in 2021 accounts for 41.7% of the total world electricity consumption, and becomes the largest world electricity consumption industry. In order to further reduce industrial energy consumption, improve the electricity utilization efficiency in the industrial field, achieve the aim and the requirements of energy conservation and emission reduction in the industrial industry, and it is important to know and optimize the electricity utilization condition in detail. Non-invasive load monitoring (NILM) is an advanced technology that can monitor its power consumption information in real time with minimal additional hardware cost and provide data support and analysis advice without affecting the operation of the various load devices of the industrial system. By introducing the NILM method in an industrial system, potential information on the health status of the device can be provided, thereby facilitating lifecycle management of the device. In recent years, the continuous progress of the artificial intelligence method in the measurement field also provides a new opportunity for the development and progress of NILM.
A typical NILM algorithm includes four steps: sampling, marking, training and deployment. The algorithm firstly samples data in the industrial system, marks the sampled data, trains a model by using the marked data, and finally deploys the trained model to monitor the real-time power consumption of each load device in the industrial system. During the training process, the NILM needs to install a sensor for each relevant load device to acquire its tag information, and after the training is finished, these sensors will be removed, only one monitoring terminal located at the main entrance is reserved, and the power consumption of each load will be identified later by using the relevant algorithm. In practice, however, this approach wastes a significant amount of data that could be used to further improve the accuracy and reliability of the model, which would further improve the recognition accuracy of the NILM algorithm if it could be better utilized.
Therefore, the present field needs to propose an algorithm structure capable of further utilizing the system to collect data based on the existing research, and further improve the accuracy and reliability of the model by deeply utilizing the existing resources.
Disclosure of Invention
In view of the above, the present invention provides a semi-supervised learning method and system for non-invasive load monitoring, which further improves the accuracy of the non-invasive load monitoring system by using the collected non-tag data in depth, and improves the reliability of the model by better using the available data in the actual scene with limited data collection.
The invention solves the problems by the following technical means:
in a first aspect, the present invention provides a semi-supervised learning method for non-invasive load monitoring, comprising a training phase for non-invasive load monitoring algorithm deployment and an reasoning phase for non-invasive load monitoring algorithm deployment;
the training phase for non-invasive load monitoring algorithm deployment comprises the following steps:
synchronously sampling the power data of the input bus and the related load of the monitored object, wherein the sampling time is set to be T seconds, and obtaining a feature set X of the monitored object 0 ＝{x 1 0 ,x 2 0 ,…,x m 0 And a load tag vector Y, where x m 0 Is a vector of length T, each element of which is a power feature, y= { Y 1 ,y 2 ,…y T Also a vector of length T, each element y T Are all in a load state, and the value range is { S } 0 ,S 1 ,…,S z S in }, S z Representing different load states, z representing the number of load states;
for the acquired data feature set X 0 Performing feature extraction processing to obtain a feature vector set X= { X 1 ,x 2 ,…，x n -wherein n represents the dimension of the feature vector;
combining the characteristic vector set X with the load label vector Y to form an initial data set D and storing the initial data set D;
after obtaining the initial dataset D, a q2seq classification model f with parameters θ is used θ () To carry out X [ t: t+L ] of length L]And Y [ t: t+L]Wherein t represents the interval start of the training dataset;
the accuracy of the model identification state result and the Mars correlation coefficient are used as screening basis, and a section with high accuracy and Ma Xiusi correlation coefficient is selected as a sampling section of the pseudo tag, so that an optimal result section is obtained;
the reasoning stage for the deployment of the non-invasive load monitoring algorithm comprises the following steps:
sampling is carried out only on data of an input bus, and the same feature extraction is carried out on the acquired data to obtain a feature vector set X= { X 1 ,x 2 ,…,x n }；
Reading a classification model obtained after initialization training, and using a seq2seq classification model f with a parameter of theta θ () For X [ t: t+L of length L]Predicting to obtain corresponding prediction results
For the prediction resultProcessing, combining the obtained optimal result interval, and using t s Extracting and storing a prediction result for each segment of the optimal result interval for step length to obtain a pseudo tag required by semi-supervised learning;
judging whether the labeling of the feature vector set X is completed or not until all new data are traversed and labeled;
after all new labels are finished, a cyclic voting mechanism is adopted, so that the accuracy of the pseudo labels is further ensured; the cyclic voting mechanism extracts labels with length of T each time in the voting implementation process, and the step length of the label is T s At the same time it must be ensured that T is T s Is a factor of N' of (c),the multiple value N' of the two is optimally adjusted according to actual conditions;
after the high-precision pseudo tag required by semi-supervised learning is obtained, the obtained high-precision pseudo tag is correspondingly grouped with input data, then is mixed with an initial data set D, and model training is carried out again, iteration is continuously repeated until a stopping standard is reached, the change of the accuracy rate of the verification set is used as the stopping standard, namely, when the accuracy rate of the verification set is reduced in the iteration process, iteration training is stopped.
Preferably, the power characteristics include one or more of the following: voltage, current, active power, reactive power, amplitude and phase angle of each subharmonic.
Preferably, the number of load states is classified by load power level.
Preferably, feature extraction is accomplished by calculating mutual information between the feature vector set X and the load tag vector Y and sorting and filtering.
Preferably, a seq2seq classification model f with a parameter θ is used θ () To carry out X [ t: t+L ] of length L]And Y [ t: t+L]Is the training of (1), the objective function is:
wherein the method comprises the steps ofIs a cross entropy loss function, seq2seq refers to a model that maps one sequence to another, L represents the length of the dataset, i represents the features in the feature vector set, and n represents the number of features contained in the feature vector set.
Preferably, the cyclic voting mechanism employs a majority voting mechanism, expressed by the following formula:
wherein the method comprises the steps ofA predicted value representing a tag employing a majority voting scheme, c representing a category selected during voting, n' being the number of judgments, +.>The label index is the label index when the model makes n' th judgment when inputting X, s represents the state result, II is an index function, and returns to 1 when the condition in brackets is satisfied, otherwise returns to 0.
Preferably, the recurring voting mechanism employs a uniform voting, expressed by the following formula:
wherein the method comprises the steps ofPredictive value, m ', N' e 1, …, N ', ∈1, N', representing a tag using a uniform vote>The label index at the m 'th and n' th judgment of the model when inputting X is respectively.
In a second aspect, the present invention provides a semi-supervised learning system for non-invasive load monitoring, comprising a training module for a training phase of a non-invasive load monitoring algorithm deployment and an inference module for an inference phase of the non-invasive load monitoring algorithm deployment;
the training module comprises:
the data synchronous sampling unit is used for synchronously sampling the power data of the input bus and the related load of the monitored object, and the sampling time is set to be T seconds to obtain a feature set X of the monitored object 0 ＝{x 1 0 ,x 2 0 ,…,x m 0 And a load tag vector Y, where x m 0 Is a vector of length T, whichEach element is a power feature, y= { Y 1 ,y 2 ,…y T Also a vector of length T, each element y T Are all in a load state, and the value range is { S } 0 ,S 1 ,…,S z S in }, S z Representing different load states, z representing the number of load states;
a data feature extraction unit for collecting data feature set X 0 Performing feature extraction processing to obtain a feature vector set X= { X 1 ,x 2 ,…x n -wherein n represents the dimension of the feature vector;
the initial data set establishing unit is used for combining the characteristic vector set X with the load label vector Y to form an initial data set D and storing the initial data set D;
a classification model training unit for employing the seq2seq classification model f with the parameter θ after obtaining the initial data set D θ () To carry out X [ t: t+L ] of length L]And Y [ t: t+L]Wherein t represents the interval start of the training dataset;
the optimal result interval obtaining unit is used for selecting an interval section with high accuracy and Ma Xiusi correlation coefficient as a sampling interval of a pseudo tag by adopting accuracy of a model identification state result and a mausk correlation coefficient as screening basis to obtain an optimal result interval;
the reasoning module comprises:
a feature vector set obtaining unit for sampling data of the input bus only, and performing the same feature extraction on the collected data to obtain a feature vector set X= { X 1 ,x 2 ,…,x n }；
A classification model prediction unit for reading the classification model obtained after the initialization training, and using the seq2seq classification model f with the parameter of θ θ () For X [ t: t+L of length L]Predicting to obtain corresponding prediction results
A prediction result extraction unit for extracting a prediction resultProcessing, combining the obtained optimal result interval, and using t s Extracting and storing a prediction result for each segment of the optimal result interval for step length to obtain a pseudo tag required by semi-supervised learning;
the judging unit is used for judging whether the labeling of the feature vector set X is completed or not until all new data are traversed and labeled;
the cyclic voting mechanism unit is used for adopting a cyclic voting mechanism after all new labels are finished, so that the accuracy of the pseudo labels is further ensured; the cyclic voting mechanism extracts labels with length of T each time in the voting implementation process, and the step length of the label is T s At the same time it must be ensured that T is T s The multiple value N' of the two is optimally adjusted according to actual conditions;
and the model retraining unit is used for mixing the obtained high-precision pseudo tag with the initial data set D after the high-precision pseudo tag required by semi-supervised learning is obtained and input data are correspondingly grouped, and retraining the model again, continuously repeating iteration until the stopping standard is reached, and taking the change of the accuracy of the verification set as the stopping standard, namely stopping iterative training when the accuracy of the verification set is reduced in the iterative process.
In a third aspect, the present invention provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the semi-supervised learning method for non-intrusive load monitoring as described in the first aspect of the invention when the program is executed.
In a fourth aspect, the present invention provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the semi-supervised learning method of non-intrusive load monitoring as described in the first aspect of this invention.
Compared with the prior art, the invention has the beneficial effects that at least:
the invention mainly comprises two stages of training and reasoning, and utilizes the prediction result of the initial training model to obtain a reliable semi-supervised learning pseudo tag through quality control, and further inputs the reliable semi-supervised learning pseudo tag into the training of the model to realize continuous self-supervised learning of the system. The method can effectively combine the unlabeled data to improve the accuracy of the model, and can better improve the accuracy and reliability of the model in the actual scene with limited data collection.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a semi-supervised learning method for non-intrusive load monitoring of the present invention;
FIG. 2 is a schematic diagram of a semi-supervised learning system for non-intrusive load monitoring of the present invention;
fig. 3 is a block diagram of the electronic device of the present invention.
Detailed Description
In order to make the above objects, features and advantages of the present invention more comprehensible, the following detailed description of the technical solution of the present invention refers to the accompanying drawings and specific embodiments. It should be noted that the described embodiments are only some embodiments of the present invention, and not all embodiments, and that all other embodiments obtained by persons skilled in the art without making creative efforts based on the embodiments in the present invention are within the protection scope of the present invention.
Example 1
The non-invasive load monitoring semi-supervised learning method adopts TCN-CRF and BITCH-CRF as the substructures, TCN and BITCH components extract features from input sequences in a model, and each time step is each possible state y t ∈{S 0 ,S 1 ,…,S z Generation of emission score p (y) t X); whereas CRF groupThe component can calculate the joint probability p (Y|X) of the output sequence on the one hand and can combine the emission score p (Y) t X) and a learned conversion score p (y) t |y t-1 ) Deducing the optimal observation sequenceWherein the conversion score p (y t |y t-1 ) Is a parameter of the CRF layer that represents the probability of a transition from one state to another.
The flow chart of the semi-supervised learning method for non-invasive load monitoring is shown in figure 1. The first part of the method is used for a training stage of non-invasive load monitoring algorithm deployment, the stage firstly synchronously samples the power data of an input bus and related loads of a monitored object, the sampling time is set to be T seconds, and a feature set X of the monitored object is obtained 0 ＝{x 1 0 ,x 2 0 ,…,x m 0 And a load tag vector Y, where x m 0 Is a vector of length T, each element of which is a power characteristic such as voltage, current, active power, reactive power, and amplitude and phase angle of each subharmonic; y= { Y 1 ,y 2 ,…y T Also a vector of length T, each element y T Are all in a load state, and the value range is { S } 0 ,S 1 ,…,S z S in }, S z Representing different load conditions, z represents the number of load conditions and can be generally classified by load power level.
To further improve the performance of the training model, the method acquires data X 0 Performing feature extraction processing to obtain a feature vector set X= { X 1 ,x 2 ,…x n Where n represents the dimension of the feature vector, feature extraction may be accomplished in an implementation by computing the mutual information between X and Y and sorting and filtering. The feature vector set X is combined with the load tag vector Y to form and store an initial data set D. After obtaining the initial dataset D, a q2seq classification model f with parameters θ is used θ () To perform X [ t ] of length L:t+L]and Y [ t: t+L]Wherein t represents the interval start of the training dataset, the objective function of which is:
wherein the method comprises the steps ofIs a cross entropy loss function, seq2seq refers to a model that maps one sequence to another, L represents the length of the dataset, i represents the features in the feature vector set, and n represents the number of features contained in the feature vector set. In the actual training process, the seq2seq model is not uniform in the whole output interval, and in order to realize the high-precision screening of the semi-supervised learning pseudo labels, the interval with higher accuracy must be screened out first to serve as the basis for generating the follow-up pseudo labels. In the process of achieving the object, the accuracy of the model identification state result and the Ma Xiusi correlation coefficient (MCC) can be used as screening basis, a section with higher accuracy and Ma Xiusi correlation coefficient is selected as a sampling section of the pseudo tag, the section is used as an optimal result section, and the follow-up semi-supervised learning pseudo tag sampling is performed in the sections.
The second part of the method is used for an reasoning stage of non-invasive load monitoring algorithm deployment, in the implementation process of the stage, sampling is only carried out on data of an input bus, and the same feature extraction is carried out on the acquired data to obtain a feature vector set X= { X 1 ,x 2 ,…,x n Then reading the model obtained after the initialization training, and using the seq2seq classification model f with the parameter of theta θ () For X [ t: t+L of length L]Predicting to obtain corresponding prediction results
In order to realize semi-supervised learning of the system, the method needs to predict the resultAnd processing to obtain the high-quality pseudo tag required by semi-supervised learning. This process requires combining the optimal result interval obtained in the first part, at t s And extracting and storing a prediction result for each segment of the optimal result interval for the step length, and judging whether the marking of the data is finished or not by using the cycle 1 shown in the figure until all new data are traversed and marked. After all data labels are marked, in order to further ensure the accuracy of pseudo labels, the invention provides a cyclic voting mechanism, and the cyclic voting mechanism extracts labels with length of T each time in the voting implementation process, and the step length of the cyclic voting mechanism is T s At the same time it must be ensured that T is T s The multiple value N' of the two can be optimally adjusted according to actual conditions, so that the accuracy of a voting mechanism is ensured; and then the data points at each time can be voted in N' estimated values to obtain a result with higher reliability. The majority voting mechanism is a basic method of generating pseudo tags, which can be represented by the following formula:
wherein the method comprises the steps ofA predicted value representing a tag employing a majority voting scheme, c representing a category selected during voting, n' being the number of judgments, +.>The label index is the label index when the model makes n' th judgment when inputting X, s represents the state result, II is an index function, and returns to 1 when the condition in brackets is satisfied, otherwise returns to 0. On the basis of this, the invention proposes a more strict mechanism than majority voting to control the effect of low-precision tags, this voting mechanism being called consistent voting and being represented by the following formula:
wherein the method comprises the steps ofPredictive value, m ', N' e 1, …, N ', ∈1, N', representing a tag using a uniform vote>The label indexes when the model is input for m 'th and n' th judgment are respectively, and the method is equivalent to only recognizing the fake label with completely consistent judgment, so that the fake label with higher accuracy required by semi-supervised learning can be obtained.
After the high-precision pseudo tag required by semi-supervised learning is obtained, the obtained pseudo tag is correspondingly grouped with input data, then is mixed with an initial data set D, and model training is carried out again, and the iteration cycle 2 is continuously repeated until a stopping standard is reached.
The invention mainly comprises two stages of training and reasoning, and utilizes the prediction result of the initial training model to obtain a reliable semi-supervised learning pseudo tag through quality control, and further inputs the reliable semi-supervised learning pseudo tag into the training of the model to realize continuous self-supervised learning of the system. The method can effectively combine the unlabeled data to improve the accuracy of the model, and can better improve the accuracy and reliability of the model in the actual scene with limited data collection.
The invention can be applied to various scenes needing non-invasive load monitoring, for example, the invention can be applied to design in the management and monitoring of a power equipment manufacturing plant, and the available data can be better utilized in the actual scene of limited data collection of the power equipment manufacturing plant so as to improve the reliability of the system. In the practical application process, the training phase of the method aims at the power load characteristic set X in the manufacturing plant 0 And a loadThe label vector Y is collected and X 0 Representing power characteristics in the manufacturing plant, such as voltage, current, active power, reactive power, and amplitude and phase angle of each subharmonic, Y represents load conditions in the manufacturing plant, and may be classified according to the load power level of the manufacturing plant. After the acquisition is finished, the power data vector set X with the load state identification and classification characteristics is obtained by carrying out feature processing on the power data vector set X, calculating mutual information between X and Y, sequencing and filtering the mutual information.
In the training process, a seq2seq classification model f with a parameter of theta is adopted θ () And training is carried out, and parameters of the classification model can be continuously debugged and set according to specific characteristics of a power equipment manufacturing plant so as to achieve better training results. In the process of obtaining the optimal result interval of the model, the accuracy of the model identification state result and Ma Xiusi related coefficient (MCC) can be used as screening basis, and in the actual application process, the optimal interval of the power equipment manufacturing plant load state prediction model is obtained by combining the analysis of the time-lapse relation of the accuracy of the power equipment manufacturing plant load state prediction, and the interval starting point and the length of the optimal interval are adjusted according to the actual value so as to ensure the reliability of the model.
In the reasoning process, the data acquisition of the power equipment manufacturing plant is only conducted on an input bus, the same characteristic extraction is carried out on the acquired data to obtain a characteristic vector set X, the power characteristics in X are the same as those extracted in the training stage, and are generally the voltage, current, active power, reactive power, amplitude, phase angle and other power characteristics of each subharmonic of the power equipment manufacturing plant, the model obtained after initialization training is further read, and a seq2seq classification model f with the parameter of theta is used θ () And predicting to obtain a load state prediction result of the power equipment manufacturing plant.
And after the load state prediction result of the power equipment manufacturing plant is obtained, analyzing and processing the load state prediction result, and further completing the generation of the pseudo tag. The process is firstly carried out at t s And extracting and storing a prediction result for each segment of the optimal result interval for the step length, and further finishing the labeling of the load running state of the power equipment. After all data is markedPseudo tag precision control is carried out on the power equipment data by adopting a cyclic voting mechanism, and each time a tag with the length of T is extracted, the step length of the tag is T s At the same time it must be ensured that T is T s The data point of each time can vote in N' estimated values to obtain a result with higher reliability. Parameter t involved in the generation of pseudo tags s The power equipment T, N' can be continuously debugged according to the actual conditions of the power equipment manufacturer so as to ensure higher accuracy and reliability.
After obtaining the high-precision pseudo tag required by the load monitoring of the power equipment manufacturing plant, the obtained pseudo tag is correspondingly grouped with input data, then mixed with an initial data set and model training is carried out again, repeated iteration is continued until a stopping standard is achieved, further continuous iteration and updating of management and monitoring model parameters of the power equipment manufacturing plant are completed, unlabeled data are converted into valuable labeled data, continuous improvement of model capability is achieved, and the method and the system are applied in the system to improve stability and reliability.
Example 2
As shown in fig. 2, the present invention provides a semi-supervised learning system for non-invasive load monitoring, comprising a training module for a training phase of a non-invasive load monitoring algorithm deployment and an inference module for an inference phase of the non-invasive load monitoring algorithm deployment;
the training module comprises a data synchronous sampling unit, a data characteristic extraction unit, an initial data set establishment unit, a classification model training unit and an optimal result interval obtaining unit;
the data synchronous sampling unit is used for synchronously sampling the power data of the input bus and the related load of the monitored object, the sampling time is set to be T seconds, and the feature set X of the monitored object is obtained 0 ＝{x 1 0 ,x 2 0 ,…,x m 0 And a load tag vector Y, where x m 0 Is a vector of length T, each element of which is a power feature, y= { Y 1 ,y 2 ,…y T Also a vector of length T, whichEach element y T Are all in a load state, and the value range is { S } 0 ,S 1 ,…,S z S in }, S z Representing different load states, z representing the number of load states;
the data feature extraction unit is used for acquiring a data feature set X 0 Performing feature extraction processing to obtain a feature vector set X= { X 1 ,x 2 ,…x n -wherein n represents the dimension of the feature vector;
the initial data set establishing unit is used for combining the characteristic vector set X with the load label vector Y to form an initial data set D and storing the initial data set D;
the classification model training unit is used for adopting a seq2seq classification model f with the parameter of theta after the initial data set D is obtained θ () To carry out X [ t: t+L ] of length L]And Y [ t: t+L]Wherein t represents the interval start of the training dataset;
the optimal result interval obtaining unit is used for selecting an interval section with high accuracy and Ma Xiusi correlation coefficient as a sampling interval of a pseudo tag by adopting accuracy of a model identification state result and a mausk correlation coefficient as screening basis to obtain an optimal result interval;
the reasoning module comprises a feature vector set obtaining unit, a classification model predicting unit, a predicting result extracting unit, a judging unit for whether all labels are marked, a cyclic voting mechanism unit and a model retraining unit;
the feature vector set obtaining unit is used for sampling data of the input bus only, and performing the same feature extraction on the acquired data to obtain a feature vector set X= { X 1 ,x 2 ,…,x n }；
The classification model prediction unit is used for reading the classification model obtained after the initialization training, and using the seq2seq classification model f with the parameter of theta θ () For X [ t: t+L of length L]Predicting to obtain corresponding prediction results
The prediction result extractionThe unit is used for predicting the resultProcessing, combining the obtained optimal result interval, and using t s Extracting and storing a prediction result for each segment of the optimal result interval for step length to obtain a pseudo tag required by semi-supervised learning;
the judging unit is used for judging whether the labeling of the feature vector set X is completed or not until all new data are traversed and labeled;
the cyclic voting mechanism unit is used for adopting a cyclic voting mechanism after all new labels are completed, so that the accuracy of the pseudo labels is further ensured; the cyclic voting mechanism extracts labels with length of T each time in the voting implementation process, and the step length of the label is T s At the same time it must be ensured that T is T s The multiple value N' of the two is optimally adjusted according to actual conditions;
the model retraining unit is used for mixing the obtained high-precision pseudo tag with the initial data set D after obtaining the high-precision pseudo tag required by semi-supervised learning and corresponding grouping of the obtained high-precision pseudo tag and the input data, and retraining the model, continuously repeating iteration until reaching a stopping standard, taking the change of the accuracy of the verification set as the stopping standard, namely stopping iterative training when the accuracy of the verification set is reduced in the iterative process.
Other features of the embodiment of the present invention are the same as those of embodiment 1, and thus are not described here again.
Example 3
Based on the same conception, the invention also provides a physical structure schematic diagram, as shown in fig. 3, the server may include: processor 810, communication interface (Communications Interface) 820, memory 830, and communication bus 840, wherein processor 810, communication interface 820, memory 830 accomplish communication with each other through communication bus 840. The processor 810 may invoke logic instructions in the memory 830 to perform the steps of the semi-supervised learning method of non-intrusive load monitoring.
Further, the logic instructions in the memory 830 described above may be implemented in the form of software functional units and may be stored in a computer-readable storage medium when sold or used as a stand-alone product. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
Example 4
Based on the same conception, the present invention also provides a non-transitory computer readable storage medium storing a computer program comprising at least one piece of code executable by a master control device to control the master control device to implement the steps of the semi-supervised learning method of non-invasive load monitoring.
In the above embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the present application are produced in whole or in part. The computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by a wired (e.g., coaxial cable, fiber optic, digital subscriber line), or wireless (e.g., infrared, wireless, microwave, etc.). The computer readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that contains an integration of one or more available media. The usable medium may be a magnetic medium (e.g., floppy disk, hard disk, tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk), etc.
Those of ordinary skill in the art will appreciate that implementing all or part of the above-described method embodiments may be accomplished by a computer program to instruct related hardware, the program may be stored in a computer readable storage medium, and the program may include the above-described method embodiments when executed. And the aforementioned storage medium includes: ROM or random access memory RAM, magnetic or optical disk, etc.
The foregoing examples illustrate only a few embodiments of the invention and are described in detail herein without thereby limiting the scope of the invention. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the invention, which are all within the scope of the invention. Accordingly, the scope of protection of the present invention is to be determined by the appended claims.
Claims (10)
1. A semi-supervised learning method for non-invasive load monitoring, comprising a training phase for non-invasive load monitoring algorithm deployment and an inference phase for non-invasive load monitoring algorithm deployment;
the training phase for non-invasive load monitoring algorithm deployment comprises the following steps:
synchronously sampling the power data of the input bus and the related load of the monitored object, wherein the sampling time is set to be T seconds, and obtaining a feature set X of the monitored object 0 ＝{x 1 0 ,x 2 0 ,…,x m 0 And a load tag vector Y, where x m 0 Is a vector of length T, each element of which is a power feature, y= { Y 1 ,y 2 ,…y T Also a vector of length T, each element y T Are all in a load state, and the value range is { S } 0 ,S 1 ,…,S z S in }, S z Representing different load states, z representing the number of load states;
for the acquired data feature set X 0 Performing feature extraction processing to obtain a feature vector set X= { X 1 ,x 2 ,…x n -wherein n represents the dimension of the feature vector;
combining the characteristic vector set X with the load label vector Y to form an initial data set D and storing the initial data set D;
after obtaining the initial dataset D, a q2seq classification model f with parameters θ is used θ () To carry out X [ t: t+L ] of length L]And Y [ t: t+L]Wherein t represents the interval start of the training dataset;
the accuracy of the model identification state result and the Mars correlation coefficient are used as screening basis, and a section with high accuracy and Ma Xiusi correlation coefficient is selected as a sampling section of the pseudo tag, so that an optimal result section is obtained;
the reasoning stage for the deployment of the non-invasive load monitoring algorithm comprises the following steps:
sampling is carried out only on data of an input bus, and the same feature extraction is carried out on the acquired data to obtain a feature vector set X= { X 1 ,x 2 ,…,x n }；
Reading a classification model obtained after initialization training, and using a seq2seq classification model f with a parameter of theta θ () For X [ t: t+L of length L]Predicting to obtain corresponding prediction results
For the prediction resultProcessing, combining the obtained optimal result interval, and using t s Extracting and storing a prediction result for each segment of the optimal result interval for step length to obtain a pseudo tag required by semi-supervised learning;
judging whether the labeling of the feature vector set X is completed or not until all new data are traversed and labeled;
after all new labels are finished, a cyclic voting mechanism is adopted, so that the accuracy of the pseudo labels is further ensured; the cyclic voting mechanism extracts labels with length of T each time in the voting implementation process, and the step length of the label is T s At the same time it must be ensured that T is T s The multiple value N' of the two is optimally adjusted according to actual conditions;
after the high-precision pseudo tag required by semi-supervised learning is obtained, the obtained high-precision pseudo tag is correspondingly grouped with input data, then is mixed with an initial data set D, and model training is carried out again, iteration is continuously repeated until a stopping standard is reached, the change of the accuracy rate of the verification set is used as the stopping standard, namely, when the accuracy rate of the verification set is reduced in the iteration process, iteration training is stopped.
2. The non-intrusive load monitoring semi-supervised learning method of claim 1, wherein the power signature comprises one or more of: voltage, current, active power, reactive power, amplitude and phase angle of each subharmonic.
3. The non-intrusive load monitoring semi-supervised learning method of claim 1, wherein the number of load conditions is categorized by load power level.
4. The semi-supervised learning method for non-intrusive load monitoring as defined in claim 1, wherein feature extraction is accomplished by computing mutual information between feature vector set X and load label vector Y and ordering and filtering.
5. The method of semi-supervised learning for non-invasive load monitoring as recited in claim 1, wherein a parameter θ, seq2seq classification model f is employed θ () To carry out X [ t: t+L ] of length L]And Y [ t: t+L]Is the training of (1), the objective function is:
wherein the method comprises the steps ofIs a cross entropy loss function, seq2seq refers to a model that maps one sequence to another, L represents the length of the dataset, i represents the features in the feature vector set, and n represents the number of features contained in the feature vector set.
6. The non-intrusive load monitoring semi-supervised learning method of claim 1, wherein the cyclic voting scheme employs a majority voting scheme represented by:
wherein the method comprises the steps ofA predicted value representing a tag employing a majority voting scheme, c representing a category selected during voting, n' being the number of judgments, +.>The label index when the model makes the n' th judgment when inputting X, s represents the state result,/and->Is a function of an index of the values of the index,return 1 when the condition in brackets is satisfied, otherwise return 0.
7. The non-intrusive load monitoring semi-supervised learning method of claim 1, wherein the recurring voting mechanism employs a uniform vote expressed by:
wherein the method comprises the steps ofPredictive value, m ', N' e 1, …, N ', ∈1, N', representing a tag using a uniform vote>The label index at the m 'th and n' th judgment of the model when inputting X is respectively.
8. A semi-supervised learning system for non-invasive load monitoring, comprising a training module for a training phase of the deployment of a non-invasive load monitoring algorithm and an inference module for an inference phase of the deployment of the non-invasive load monitoring algorithm;
the training module comprises:
the data synchronous sampling unit is used for synchronously sampling the power data of the input bus and the related load of the monitored object, and the sampling time is set to be T seconds to obtain a feature set X of the monitored object 0 ＝{x 1 0 ,x 2 0 ,…,x m 0 And a load tag vector Y, where x m 0 Is a vector of length T, each element of which is a power feature, y= { Y 1 ,y 2 ,…y T Also a vector of length T, each element y T Are all in a load state, and the value range is { S } 0 ,S 1 ,…,S z S in }, S z Representing different load states, z representing the number of load states;
a data feature extraction unit for collecting data feature set X 0 Performing feature extraction processing to obtain a feature vector set X= { X 1 ,x 2 ,…x n -wherein n represents the dimension of the feature vector;
the initial data set establishing unit is used for combining the characteristic vector set X with the load label vector Y to form an initial data set D and storing the initial data set D;
a classification model training unit for employing the seq2seq classification model f with the parameter θ after obtaining the initial data set D θ () To carry out X [ t: t+L ] of length L]And Y [ t: t+L]Wherein t represents the interval start of the training dataset;
the optimal result interval obtaining unit is used for selecting an interval section with high accuracy and Ma Xiusi correlation coefficient as a sampling interval of a pseudo tag by adopting accuracy of a model identification state result and a mausk correlation coefficient as screening basis to obtain an optimal result interval;
the reasoning module comprises:
a feature vector set obtaining unit for sampling data of the input bus only, and performing the same feature extraction on the collected data to obtain a feature vector set X= { X 1 ,x 2 ,…,x n }；
A classification model prediction unit for reading the classification model obtained after the initialization training, and using the seq2seq classification model f with the parameter of θ θ () For X [ t: t+L of length L]Predicting to obtain corresponding prediction results
A prediction result extraction unit for extracting a prediction resultProcessing, combining the obtained optimal result interval, and using t s Extracting and storing the prediction result of each segment of the optimal result interval for step length to obtainPseudo tags required for semi-supervised learning;
the judging unit is used for judging whether the labeling of the feature vector set X is completed or not until all new data are traversed and labeled;
the cyclic voting mechanism unit is used for adopting a cyclic voting mechanism after all new labels are finished, so that the accuracy of the pseudo labels is further ensured; the cyclic voting mechanism extracts labels with length of T each time in the voting implementation process, and the step length of the label is T s At the same time it must be ensured that T is T s The multiple value N' of the two is optimally adjusted according to actual conditions;
and the model retraining unit is used for mixing the obtained high-precision pseudo tag with the initial data set D after the high-precision pseudo tag required by semi-supervised learning is obtained and input data are correspondingly grouped, and retraining the model again, continuously repeating iteration until the stopping standard is reached, and taking the change of the accuracy of the verification set as the stopping standard, namely stopping iterative training when the accuracy of the verification set is reduced in the iterative process.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the semi-supervised learning method for non-invasive load monitoring as recited in any of claims 1-7 when the program is executed by the processor.
10. A non-transitory computer readable storage medium, having stored thereon a computer program, which when executed by a processor, implements the steps of the semi-supervised learning method for non-intrusive load monitoring as defined in any of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311426481.8A CN117473368A (en) | 2023-10-30 | 2023-10-30 | Non-invasive load monitoring semi-supervised learning method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311426481.8A CN117473368A (en) | 2023-10-30 | 2023-10-30 | Non-invasive load monitoring semi-supervised learning method and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117473368A true CN117473368A (en) | 2024-01-30 |
Family
ID=89626877
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311426481.8A Withdrawn CN117473368A (en) | 2023-10-30 | 2023-10-30 | Non-invasive load monitoring semi-supervised learning method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117473368A (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110376457A (en) * | 2019-06-28 | 2019-10-25 | 同济大学 | Non-intrusion type load monitoring method and device based on semi-supervised learning algorithm |
| CN112712203A (en) * | 2020-12-29 | 2021-04-27 | 湖南大学 | Method and system for predicting daily maximum load of power distribution network |
| CN113595242A (en) * | 2021-06-28 | 2021-11-02 | 湖南大学 | Non-invasive load identification method based on deep CNN-HMM |
| 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 |
-
2023
- 2023-10-30 CN CN202311426481.8A patent/CN117473368A/en not_active Withdrawn
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110376457A (en) * | 2019-06-28 | 2019-10-25 | 同济大学 | Non-intrusion type load monitoring method and device based on semi-supervised learning algorithm |
| CN112712203A (en) * | 2020-12-29 | 2021-04-27 | 湖南大学 | Method and system for predicting daily maximum load of power distribution network |
| CN113595242A (en) * | 2021-06-28 | 2021-11-02 | 湖南大学 | Non-invasive load identification method based on deep CNN-HMM |
| 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 |
Non-Patent Citations (1)
| Title |
|---|
| ZHENYU ZHANG ET AL: "Achieving Sustained Improvement inIdentification Accuracy with a Semi-Supervised Learning Approach for NILM Systems", 《TECHRXIV》, 7 April 2023 (2023-04-07), pages 2 - 5 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110596506A (en) | Converter fault diagnosis method based on time convolution network | |
| CN112215405A (en) | Non-invasive type residential electricity load decomposition method based on DANN domain adaptive learning | |
| CN112215406A (en) | Non-invasive type residential electricity load decomposition method based on time convolution neural network | |
| CN112420055A (en) | Substation state identification method and device based on voiceprint characteristics | |
| CN112305388B (en) | On-line monitoring and diagnosing method for insulation partial discharge faults of generator stator winding | |
| CN112529678A (en) | Financial index time sequence abnormity detection method based on self-supervision discriminant network | |
| CN118100171B (en) | Power prediction method and system of energy conversion system and electronic equipment | |
| CN116503354A (en) | Method and device for detecting and evaluating hot spots of photovoltaic cells based on multi-mode fusion | |
| CN116721458A (en) | Cross-modal time sequence contrast learning-based self-supervision action recognition method | |
| CN111949707A (en) | Shadow field-based hidden Markov model non-invasive load decomposition method | |
| CN117056865B (en) | Method and device for diagnosing operation faults of machine pump equipment based on feature fusion | |
| CN112836755B (en) | Sample image generation method and system based on deep learning | |
| CN117942038A (en) | Sleep stage identification system integrating cross prediction and discrimination tasks | |
| CN117669656A (en) | TCN-Semi PN-based direct-current micro-grid stability real-time monitoring method and device | |
| CN117596191A (en) | Power Internet of things abnormality detection method, device and storage medium | |
| CN117079005A (en) | Optical cable fault monitoring method, system, device and readable storage medium | |
| CN117473368A (en) | Non-invasive load monitoring semi-supervised learning method and system | |
| CN109934248B (en) | Multi-model random generation and dynamic self-adaptive combination method for transfer learning | |
| CN116992757A (en) | Wellhead pressure prediction method and device based on deep learning and rolling optimization | |
| CN116028893A (en) | Intelligent data monitoring method, system and electronic equipment | |
| CN110866840A (en) | Database modeling method for power load characteristic quantity training based on knowledge graph | |
| CN109886109B (en) | Behavior identification method based on deep learning | |
| CN110033082B (en) | Method for identifying deep learning model in AI (Artificial intelligence) equipment | |
| CN112668713A (en) | Intelligent energy consumption prediction method and system | |
| CN115063643B (en) | Automatic labeling method for power scene image data |
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 |
Application publication date: 20240130 |
|
| WW01 | Invention patent application withdrawn after publication |