Load feature extraction method, load identification method and system based on multi-label learning
By adopting multi-label learning and classifier chain model in load recognition, load characteristics are converted into binary feature tags, the problem of label imbalance is solved, the accuracy of load recognition is improved, and it is suitable for load recognition of diverse electrical equipment.
Patent Information
- Application Number
- CN202411727249.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-28
- Publication Date
- 2025-03-21
AI Technical Summary
In the existing load identification method based on supervised learning model, there is a problem of label imbalance, which leads to the low prediction accuracy of labels for smaller comparisons of the model, making it difficult to cope with the load identification needs of diversified electrical equipment.
The load feature extraction method based on multi-label learning is adopted, and the load feature is converted into binary feature labels through the classifier chain model, and the prediction accuracy of the model is improved through multi-label training.
It effectively solves the problem of label imbalance, improves the accuracy of load identification, and can better respond to the load identification needs of diversified electrical equipment.
Smart Images
Figure CN119669915A_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the technical field of load identification, and specifically relates to a load feature extraction method, a load identification method and a system based on multi-label learning. Background Art
[0002] Load identification is one of the key technologies in the power system and is widely used in power load monitoring, energy efficiency management, smart grid and other fields. Its goal is to identify the operating status and power consumption of different types of electrical equipment through real-time monitoring and analysis of power load, so as to provide data support for power dispatching, demand response and energy management. At present, supervised learning models are widely used in load identification. Labels are used in supervised learning to provide target outputs for each sample, helping the model to achieve prediction by learning the relationship between input and output. At present, most of the load identification methods based on supervised learning models have the problem of unbalanced label categories, that is, the proportion of samples with different labels in the training sample set is quite different, resulting in low prediction accuracy of the model for labels with a smaller proportion, which makes it difficult to cope with the load identification needs of diversified electrical equipment in practice. Summary of the invention
[0003] The purpose of this application is to provide a load feature extraction method, a load identification method and a system based on multi-label learning to solve the label imbalance problem in load identification.
[0004] To achieve the above-mentioned purpose, the first aspect of the present application provides a load feature extraction method based on multi-label learning, which is: using a trained classifier chain model to predict the binary feature label of an input feature vector; wherein the feature vector is composed of load features extracted from power consumption information, including power load features; the classifier chain model includes multiple types of labels and binary classifiers corresponding to each type of label, and the labels are sorted in the classifier chain, and one type of label corresponds to one type of electrical equipment;
[0005] The training of the classifier chain model includes:
[0006] S100: construct training sample sets for each type of label respectively, the training sample sets including feature vectors and binary label values of each training sample; when the device type label of the training sample corresponds to the current label, its binary label value is 1; otherwise, it is 0;
[0007] S200: using the training sample set to train the binary classifiers corresponding to the labels respectively;
[0008] S300: performing multi-label training on each binary classifier in sequence according to the order of the labels in the classifier chain;
[0009] This step further includes:
[0010] S310: Initialize label number j=2;
[0011] S320: Use the binary classifier of the j-1th class label to predict the test set X * The multi-label component Using multi-label training sets Train a binary classifier for the jth class label;
[0012] S330: Let j=j+1, and execute step S320 again until multi-label training is completed for all binary classifiers.
[0013] In some embodiments, the feature vector also includes non-electrical load features.
[0014] In some specific embodiments, the binary classifier is constructed by selecting a neural network model or a decision tree classification model.
[0015] In some specific implementations, before constructing the training sample sets of various labels, an initial sample set is first constructed, the initial sample set is divided into an initial training set and a test set, and then the initial training set is converted into the training sample sets of various labels;
[0016] The above initial sample set is constructed, including:
[0017] Collect electricity consumption information of different types of electrical equipment respectively, extract load characteristics from the electricity consumption information and form feature vectors, and manually annotate equipment type labels;
[0018] The above converts the initial training set into training sample sets of various labels, including:
[0019] The training sample sets of various labels are represented as Where j is the label number in the classifier chain, i is the training sample number, and n is the number of training samples; X represents the feature vector set of the training sample, and the feature vector data in the initial training set is taken; y j Represents the j-th label component of the training sample, which is composed of the binary label values of each training sample.
[0020] The second aspect of the present application provides a load identification method based on multi-label learning, comprising:
[0021] Extract load characteristics from power consumption information to form a feature vector;
[0022] Use the trained classifier chain model to predict the binary feature labels of the input feature vector;
[0023] The load identification model is used to identify the binary feature labels;
[0024] Among them, the classifier chain model includes multiple types of labels and binary classifiers corresponding to each type of label. The labels are sorted in the classifier chain, and one type of label corresponds to one type of electrical equipment;
[0025] The training of the above classifier chain model includes:
[0026] S100: construct training sample sets for each type of label respectively, the training sample sets including feature vectors and binary label values of each training sample; when the device type label of the training sample corresponds to the current label, its binary label value is 1; otherwise, it is 0;
[0027] S200: using the training sample set to train the binary classifiers corresponding to the labels respectively;
[0028] S300: performing multi-label training on each binary classifier in sequence according to the order of the labels in the classifier chain;
[0029] This step further includes:
[0030] S310: Initialize label number j=2;
[0031] S320: Use the binary classifier of the j-1th class label to predict the test set X * The multi-label component Using multi-label training sets Train a binary classifier for the jth class label;
[0032] S330: Let j=j+1, and execute step S320 again until multi-label training is completed for all binary classifiers.
[0033] In some specific embodiments, the load identification model selects a support vector machine model or a random forest model.
[0034] The third aspect of the present application provides a load feature extraction system based on multi-label learning, including a load feature extraction module and a training module;
[0035] A load feature extraction module, which is used to predict the binary feature label of the input feature vector using the trained classifier chain model; wherein the feature vector is composed of load features extracted from the power consumption information, including power load features; the classifier chain model includes multiple types of labels and binary classifiers corresponding to each type of label, and the labels are sorted in the classifier chain, and one type of label corresponds to one type of electrical equipment;
[0036] A training module, which is used to train the classifier chain model;
[0037] The training module further includes:
[0038] The training sample set construction submodule is used to construct training sample sets of various labels. The training sample sets include the feature vector and binary label value of each training sample. When the device type label of the training sample corresponds to the current label, its binary label value is 1; otherwise, it is 0.
[0039] The training submodule is used to train the binary classifiers of the corresponding labels using the training sample sets;
[0040] The multi-label training submodule is used to perform multi-label training on each binary classifier in sequence according to the order of the labels in the classifier chain; the multi-label training submodule is configured as follows:
[0041] S310: Initialize label number j=2;
[0042] S320: Use the binary classifier of the j-1th class label to predict the test set X * The multi-label component Using multi-label training sets Train a binary classifier for the jth class label;
[0043] S330: Let j=j+1, and execute step S320 again until multi-label training is completed for all binary classifiers.
[0044] A fourth aspect of the present application provides a load identification method based on multi-label learning, comprising a feature vector construction module, a load feature extraction module, a load identification module and a training module;
[0045] A feature vector construction module is used to extract load features from power consumption information to form a feature vector;
[0046] A load feature extraction module, which is used to predict the binary feature label of the input feature vector using the trained classifier chain model; wherein the feature vector is composed of load features extracted from the power consumption information, including power load features; the classifier chain model includes multiple types of labels and binary classifiers corresponding to each type of label, and the labels are sorted in the classifier chain, and one type of label corresponds to one type of electrical equipment;
[0047] A load identification module, which is used to perform load identification on binary feature labels using a load identification model;
[0048] A training module, which is used to train the classifier chain model; the training module further includes:
[0049] The training sample set construction submodule is used to construct training sample sets of various labels. The training sample sets include the feature vector and binary label value of each training sample. When the device type label of the training sample corresponds to the current label, its binary label value is 1; otherwise, it is 0.
[0050] The training submodule is used to train the binary classifiers of the corresponding labels using the training sample sets;
[0051] The multi-label training submodule is used to perform multi-label training on each binary classifier in sequence according to the order of the labels in the classifier chain; the multi-label training submodule is configured as follows:
[0052] S310: Initialize label number j=2;
[0053] S320: Use the binary classifier of the j-1th class label to predict the test set X * The multi-label component Using multi-label training sets Train a binary classifier for the jth class label;
[0054] S330: Let j=j+1, and execute step S320 again until multi-label training is completed for all binary classifiers.
[0055] In some specific embodiments, the load identification model selects a support vector machine model or a random forest model.
[0056] Compared with the prior art, this application has the following advantages and beneficial effects:
[0057] 1. This application can solve the problem of label imbalance in load identification and can meet the load identification needs of diversified electrical equipment in practice;
[0058] 2. This application utilizes a classifier chain model to convert load features into binary feature labels associated with the load features. The binary feature labels contain more load feature information and are applied to load identification to enhance the accuracy of load identification. BRIEF DESCRIPTION OF THE DRAWINGS
[0059] Figure 1 A schematic diagram of a flow chart of a load feature extraction method provided in Example 1;
[0060] Figure 2 Schematic diagram of the training process of the classifier chain model in Example 1;
[0061] Figure 3 A flow chart of the load identification method provided in Example 2. DETAILED DESCRIPTION
[0062] The following will be combined with the accompanying drawings to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0063] Example 1
[0064] Please refer to Figure 1 , which is a flow chart of the load feature extraction method provided by this embodiment. In this embodiment, the load feature extraction method based on multi-label learning is: using a trained classifier chain model to predict the binary feature label of the input feature vector, that is, the load feature based on multi-label learning; wherein the feature vector is composed of load features extracted from power consumption information; the power consumption information includes current, voltage, active power, reactive power and other data. In this embodiment, the power consumption information is obtained in a non-invasive manner.
[0065] Extracting load characteristics from electricity consumption information is a mature technology in the industry, and its specific implementation process will not be repeated here. Load characteristics include power load characteristics and non-power load characteristics. Common power load characteristics include but are not limited to power peak, power valley, start and stop times, harmonic characteristics, and common non-power load characteristics include but are not limited to usage time. In this embodiment, several load characteristics are pre-selected, and the selected load characteristics are extracted by analyzing the electricity consumption information to construct a feature vector.
[0066] The classifier chain model includes multiple categories of labels and binary classifiers corresponding to each category of labels. The labels are sorted in the classifier chain, and one category of labels represents a category of electrical equipment. A separate binary classifier is trained for each label. The binary classifier is a mapping function model from feature vectors to binary label values. Therefore, the binary classifier can be used to predict the binary label value corresponding to the instance.
[0067] The basic unit of the classifier chain model is the binary classifier, which is a mapping function model from feature vector to binary label value. The binary classifier can be constructed by a neural network model or a decision tree classification model.
[0068] The idea of a binary classifier is as follows:
[0069] f:χ→2 y (1)
[0070] In formula (1), χ represents the input feature vector; y = {λ1,λ2,...,λ q}, y is the label space of class q, λ j Represents the j-th class label vector, j = 1, 2, ... q.
[0071] Please refer to Figure 2 , which is a schematic diagram of the training process of the classifier chain model in this embodiment, will be combined with Figure 2 Describe in detail the training process of the classifier chain model.
[0072] S100: construct training sample sets for each type of label respectively, the training sample sets including feature vectors and binary label values of each training sample; when the device type label of the training sample corresponds to the current label, its binary label value is 1; otherwise, it is 0;
[0073] In this embodiment, the equipment type labels of the training samples are manually annotated. Since one type of label in the classifier chain model corresponds to one type of electrical equipment, there is a corresponding relationship between the equipment type label and the label in the classifier chain model. By judging whether the equipment type label corresponds to the current label, the binary label value of the training sample is determined.
[0074] For example, when constructing a training sample set of the j-th label, it is determined whether the device type label of the training sample corresponds to the j-th label. If so, the binary label value of the training sample is 1; otherwise, it is 0.
[0075] In this embodiment, before constructing the training sample sets of various labels, an initial sample set is first constructed, the initial sample set is divided into an initial training set and a test set, and then the initial training set is converted into the training sample sets of various labels.
[0076] Specifically, constructing an initial sample set includes:
[0077] Collect electricity consumption information of different types of electrical equipment, extract load characteristics from the electricity consumption information and form feature vectors, and manually annotate the equipment type labels. For example, for the feature vector extracted from the electricity consumption information of a kettle, its equipment type label is annotated as a kettle.
[0078] In this embodiment, the initial training set can be expressed as The number of training samples is n. For the i-th training sample (x (i) ,y (i) ), x (i) represents the feature vector of the i-th training sample, y (i) Represents the device type label of the i-th training sample.
[0079] Specifically, the initial training set is converted into training sample sets of various labels based on binary correlation, including:
[0080] The training sample sets of various labels are represented as Where j is the label number in the classifier chain, i is the training sample number, and n is the number of training samples; X represents the feature vector set of the training sample, and the feature vector data in the initial training set is taken; y j Represents the j-th label component of the training sample, which is composed of the binary label values of each training sample.
[0081] The feature vector set of the training sample is represented as X = [x (1) ,x(2) ,...x (n) ] T , where x (i) represents the feature vector of the i-th training sample; the j-th class label component of the training sample is represented by y j =[y j (1) ,y j (2) ,...y j (n) ] T , where y j (i) represents the binary label value of the i-th training sample. When the device type label of the i-th training sample corresponds to the j-th class label in the classifier chain, then y j (i) Take 1; otherwise y j (i) Take 0.
[0082] S200: using the training sample set to train the binary classifiers of the corresponding labels respectively.
[0083] The binary classifier trained this time takes the corresponding label component as output. For example, the binary classifier corresponding to the j-th label is represented as g j :X→y j , which takes the feature vector as input and the j-th class label component as output.
[0084] S300: performing multi-label training on each binary classifier in sequence according to the order of the labels in the classifier chain;
[0085] This step further includes:
[0086] S310: Initialize label number j=2;
[0087] S320: Use the binary classifier of the j-1th class label to predict the test set X * The multi-label component Using multi-label training sets Perform multi-label training on the binary classifier of the j-th label;
[0088] When performing multi-label training on a binary classifier for the jth label, the multi-label training set is used as input and the multi-label component set is used as input. is the output.
[0089] S330: Let j=j+1, and execute step S320 again until multi-label training is completed for all binary classifiers.
[0090] In the multi-label training of each binary classifier, the test set X is used *Can be the same or different.
[0091] When the trained classifier chain model is used to predict the binary feature label of an instance, the instance is input into the trained classifier chain model, and the multi-label component set predicted by the previous binary classifier and the input instance are used as the output of the next binary classifier. For the binary classifier of the jth class label, the predicted multi-label component set is y = [y1, y2, ... y j ] is the output, y1, y2, ...y j Respectively represent the 1st, 2nd, ...jth class label components of the instance; and so on, the last binary classifier outputs the predicted binary feature label y = [y1, y2, ...y d ], d represents the number of labels, and the label weight of the instance under all labels can be obtained.
[0092] Example 2
[0093] This embodiment provides a load identification method, including: obtaining the user's total electricity consumption information in a non-invasive manner, extracting load features from the electricity consumption information to form a feature vector; using a trained classifier chain model to predict the binary feature label of the input feature vector; using a load identification model to perform load identification on the binary feature label; in this embodiment, the load identification model selects a deep learning model, such as a support vector machine model or a random forest model.
[0094] The present application utilizes a classifier chain model to convert load characteristics into binary feature labels associated with the load characteristics. The binary feature labels contain more load characteristic information and are applied to load identification to enhance the accuracy of load identification.
[0095] Note that the above are only preferred embodiments of the present application and the technical principles used. Those skilled in the art will understand that the present application is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the protection scope of the present application. Therefore, although the present application is described in more detail through the above embodiments, the present application is not limited to the above embodiments, and may also include more other equivalent embodiments without departing from the concept of the present application, all of which belong to the protection scope of the present application.
Claims
1. Load feature extraction method based on multi-label learning, characterized by: Use the trained classifier chain model to predict the binary feature labels of the input feature vector; The feature vector is composed of load features extracted from the power consumption information, including power load features; The classifier chain model includes multiple types of labels and binary classifiers corresponding to each type of label. The labels are sorted in the classifier chain, and one type of label corresponds to one type of electrical equipment. The training of the classifier chain model includes: S100: construct training sample sets for each type of label respectively, the training sample sets including feature vectors and binary label values of each training sample; when the device type label of the training sample corresponds to the current label, its binary label value is 1; otherwise, it is 0; S200: using the training sample set to train the binary classifiers corresponding to the labels respectively; S300: performing multi-label training on each binary classifier in sequence according to the order of the labels in the classifier chain; This step further includes: S310: Initialize label number j=2; S320: Use the binary classifier of the j-1th class label to predict the test set X * The multi-label component Using multi-label training sets Train a binary classifier for the jth class label; S330: Let j=j+1, and execute step S320 again until multi-label training is completed for all binary classifiers.
2. The load feature extraction method according to claim 1, characterized in that: The feature vector also includes non-electrical load features.
3. The load feature extraction method according to claim 1, characterized in that: The binary classifier is constructed by selecting a neural network model or a decision tree classification model.
4. The load feature extraction method according to claim 1, characterized in that: Before constructing the training sample sets of various labels, an initial sample set is first constructed, the initial sample set is divided into an initial training set and a test set, and then the initial training set is converted into the training sample sets of various labels; The constructing of the initial sample set comprises: Collect electricity consumption information of different types of electrical equipment respectively, extract load characteristics from the electricity consumption information and form feature vectors, and manually annotate equipment type labels; The initial training set is converted into a training sample set of various labels, including: The training sample sets of various labels are represented as Where j is the label number in the classifier chain, i is the training sample number, and n is the number of training samples; X represents the feature vector set of the training sample, and the feature vector data in the initial training set is taken; y j Represents the j-th label component of the training sample, which is composed of the binary label values of each training sample.
5. The load identification method based on multi-label learning is characterized by: include: Extract load characteristics from power consumption information to form a feature vector; Use the trained classifier chain model to predict the binary feature labels of the input feature vector; The load identification model is used to identify the binary feature labels; The classifier chain model includes multiple types of labels and binary classifiers corresponding to each type of label. The labels are sorted in the classifier chain, and one type of label corresponds to one type of electrical equipment. The training of the classifier chain model includes: S100: construct training sample sets for each type of label respectively, the training sample sets including feature vectors and binary label values of each training sample; when the device type label of the training sample corresponds to the current label, its binary label value is 1; otherwise, it is 0; S200: using the training sample set to train the binary classifiers corresponding to the labels respectively; S300: performing multi-label training on each binary classifier in sequence according to the order of the labels in the classifier chain; This step further includes: S310: Initialize label number j=2; S320: Use the binary classifier of the j-1th class label to predict the test set X * The multi-label component Using multi-label training sets Train a binary classifier for the jth class label; S330: Let j=j+1, and execute step S320 again until multi-label training is completed for all binary classifiers.
6. The load identification method according to claim 1, characterized in that: The load identification model selects a support vector machine model or a random forest model.
7. Load feature extraction system based on multi-label learning, characterized by: Including load feature extraction module and training module; The load feature extraction module is used to predict the binary feature label of the input feature vector using the trained classifier chain model; The feature vector is composed of load features extracted from power consumption information, including power load features; the classifier chain model includes multiple types of labels and binary classifiers corresponding to each type of label, and the labels are sorted in the classifier chain, and one type of label corresponds to one type of electrical equipment; The training module is used to train the classifier chain model; wherein the training module further includes: The training sample set construction submodule is used to construct training sample sets of various labels. The training sample sets include the feature vector and binary label value of each training sample. When the device type label of the training sample corresponds to the current label, its binary label value is 1; otherwise, it is 0. The training submodule is used to train the binary classifiers of the corresponding labels using the training sample sets; The multi-label training submodule is used to perform multi-label training on each binary classifier in sequence according to the order of the labels in the classifier chain; The multi-label training submodule is configured as follows: S310: Initialize label number j=2; S320: Use the binary classifier of the j-1th class label to predict the test set X * The multi-label component Using multi-label training sets Train a binary classifier for the jth class label; S330: Let j=j+1, and execute step S320 again until multi-label training is completed for all binary classifiers.
8. A load identification method based on multi-label learning, characterized by: It includes a feature vector construction module, a load feature extraction module, a load identification module and a training module; The feature vector construction module is used to extract load features from power consumption information to form a feature vector; The load feature extraction module is used to predict the binary feature label of the input feature vector using the trained classifier chain model; The feature vector is composed of load features extracted from power consumption information, including power load features; the classifier chain model includes multiple types of labels and binary classifiers corresponding to each type of label, and the labels are sorted in the classifier chain, and one type of label corresponds to one type of electrical equipment; The load identification module is used to perform load identification on binary feature tags using a load identification model; The training module is used to train the classifier chain model; wherein the training module further includes: The training sample set construction submodule is used to construct training sample sets of various labels. The training sample sets include the feature vector and binary label value of each training sample. When the device type label of the training sample corresponds to the current label, its binary label value is 1; otherwise, it is 0. The training submodule is used to train the binary classifiers of the corresponding labels using the training sample sets; The multi-label training submodule is used to perform multi-label training on each binary classifier in sequence according to the order of the labels in the classifier chain; The multi-label training submodule is configured as follows: S310: Initialize label number j=2; S320: Use the binary classifier of the j-1th class label to predict the test set X * The multi-label component Using multi-label training sets Train a binary classifier for the jth class label; S330: Let j=j+1, and execute step S320 again until multi-label training is completed for all binary classifiers.
9. The load identification method according to claim 8, characterized in that: The load identification model selects a support vector machine model or a random forest model.