Power usage behavior pattern recognition method and apparatus based on noisy learning

Abstract

The application discloses a power use behavior mode recognition method and device based on noisy learning, and the method comprises the following steps: through an Adjmatch-based noisy learning method, daily power consumption record data of a family is screened, the daily power consumption record data is divided into two parts of clean data and noisy data, a semi-supervised mode is used for training, category prior information is introduced to correct a loss function, and based on a contrast learning technology, two contrast loss functions are used to enhance the mode recognition capability of a neural network on daily power consumption record data with noise labels. Through analysis on the daily power consumption record of the family, the corresponding power consumption mode is recognized. Compared with a conventional power use behavior mode recognition method based on deep learning, the negative effect of noise labels on recognition accuracy can be effectively reduced, so that the accuracy of mode recognition is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to power usage behavior pattern recognition, in particular to a power usage behavior pattern recognition method and device based on noisy learning. BACKGROUND

[0002] With the development of social economy, energy consumption is increasing, and energy shortage and environmental pollution have become important problems restricting human development. Therefore, how to reasonably use and manage energy and improve energy utilization efficiency has become a problem to be solved at present. Power usage behavior pattern recognition technology can help us manage power usage more finely and improve energy utilization efficiency. With the development of emerging technologies such as smart grid, the identification of power usage behavior patterns becomes more and more important.

[0003] At present, methods based on machine learning have been widely used to identify power usage behavior, but due to the complexity of power usage behavior patterns and the problem of noisy data labels, the accuracy and robustness of these methods still face certain challenges. Noise is defined as any interruption in the observed relationship between the features of an instance and its class in time series data. Noise can occur in various ways, such as incorrect labeling by human annotators, measurement errors or natural changes in data, which makes it difficult to analyze power usage behavior patterns from power usage record data.

[0004] For the above reasons, the present application proposes a power usage behavior pattern recognition method and device based on noisy learning to improve the accuracy and robustness of power usage behavior pattern recognition. SUMMARY

[0005] The purpose of the present application is to overcome the defects of existing power usage behavior pattern recognition methods based on machine learning when the data labels are noisy. The present application provides a power usage behavior pattern recognition method and device based on noisy learning, which uses a simple data filtering method to select power usage record data with clean labels, then treats the remaining data as unlabeled data, and uses a semi-supervised method to train the neural network, so that the neural network learns discriminative behavior pattern features from daily power usage record data, and is robust to data with noisy labels.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] A power usage behavior pattern recognition method based on noisy learning, comprising the following steps:

[0008] A data collection step, collecting daily power usage record data of a household, labeling the data according to its usage behavior pattern, and forming a training data set;

[0009] The data preprocessing step preprocesses the training set data, including removing outliers, filling missing values, normalizing, and data augmentation, which includes weak augmentation and strong augmentation;

[0010] The model preheating step inputs the daily power consumption records in the training data set into two neural networks without distinction for preliminary training, so that the neural networks have preliminary recognition ability for power usage behavior patterns;

[0011] The data screening step uses a Gaussian mixture model to screen the power consumption record data, divides the data into two parts, each with clean labels and noise labels, and forms a labeled set and an unlabeled set;

[0012] The model training step uses the divided daily power consumption record data for model training, calculates the corrected cross-entropy loss using labeled data, calculates the consistency regularization loss using unlabeled data, and obtains refined behavior pattern features using contrastive learning;

[0013] The pattern recognition step uses the trained model to analyze the power consumption record data to obtain the recognition result of the power usage behavior pattern.

[0014] In the data screening step, two-part Gaussian mixture models are used to model the loss distribution of daily power consumption record data obtained through neural networks. The EM algorithm is used to solve the parameters of the Gaussian mixture model, and then the probability of each power consumption record data belonging to each Gaussian part is calculated. According to the set screening threshold, clean data is screened as labeled data, and the remaining power consumption record data is used as unlabeled data.

[0015] In the training step, due to the inevitable division of some noise data into the labeled set during the data screening stage, the noise data will be gradually overfitted during training. Therefore, in order to avoid the confidence of the incorrect recognition result of the power usage behavior pattern of the noise data being too large, the co-teaching training method is used, and two neural networks are trained collaboratively. The time series selected by one neural network is used for training by another neural network. Since the parameters are different, the above problem can be effectively avoided.

[0016] In the training step, due to the inconsistent number of classes in the labeled data obtained through the data screening module, the classifier has a bias, so the cross-entropy loss is corrected by introducing class information to obtain a supervised cross-entropy loss corrected by class:

[0017]

[0018] where L is the labeled data set, πy is an estimate of the prior , tau is a hyperparameter, tau > 0, f y (x) represents the logit output predicted by the neural network for the sample x to correspond to the label y, and y′ represents the sum of all label categories.

[0019] In the training step, in order to make full use of the power consumption record data, a consistency regularization loss is introduced to train the unlabeled data, and the specific loss function form is as follows:

[0020]

[0021] Where H(p||q) represents the cross-entropy loss of two distributions, alpha(x) represents the data augmentation copy obtained by weak enhancement, represents the data augmentation copy obtained by strong enhancement, q θ (·) represents the prediction result of the neural network output, and T represents the temperature coefficient.

[0022] In the training step, considering that there is data with unobvious power consumption behavior pattern characteristics, by introducing contrast learning, the transformation consistency of the power consumption record data is strengthened, and specifically, two kinds of contrast loss functions are included, including space-time contrast loss and instance contrast loss, so that the neural network has better recognition ability for power consumption behavior patterns. The form of the contrast loss function is as follows:

[0023]

[0024] Where alpha is a hyperparameter, representing the proportion of l inst and l temp .

[0025] The application also provides a novel power consumption behavior pattern recognition device, comprising:

[0026] A data acquisition module is used to collect daily power consumption record data of a family, and according to the use behavior pattern, the data is labeled to form a training data set;

[0027] A data preprocessing module is used to preprocess the training set data, including removing outliers, filling missing values, normalizing, and data augmentation, wherein the data augmentation includes two modes of weak enhancement and strong enhancement;

[0028] A model preheating module is used to input the daily power consumption record in the training data set into two neural networks without distinction for preliminary training, so that the neural network has a preliminary recognition ability for the power consumption behavior pattern;

[0029] The data screening module is used for screening the power consumption record data by using a Gaussian mixture model, and dividing the data into two parts, respectively with clean labels and with noise labels, to form a labeled set and an unlabeled set.

[0030] The model training module is used for model training by using the divided daily power consumption record data, calculating a corrected cross-entropy loss by using the labeled data, calculating a consistency regularization loss by using the unlabeled data, and obtaining refined behavior pattern features by using contrastive learning.

[0031] The pattern recognition module is used for analyzing the power consumption record data according to the trained model to obtain a recognition result of the power usage behavior pattern.

[0032] Beneficial effects: (1) The present application can identify the corresponding power usage behavior pattern according to the daily power consumption record data of the family, and can effectively detect and screen the data with noise labels without relying on a noise model, and has strong recognition ability for the power usage behavior pattern. (2) The present application uses a Gaussian mixture model (GMM) to screen the daily power consumption record data, screens out the data with clean labels as labeled data for training, and uses a correction for cross-entropy loss to avoid the influence of class imbalance. The remaining power consumption record is trained by using a consistency regularization-based method, and the data is fully utilized. In addition, considering that there is data in the time series that is difficult to analyze the power usage behavior pattern, two contrastive losses are introduced to further improve the recognition ability of the neural network for the power usage behavior pattern. (3) The present application collects actual daily power consumption record data of the family for training and testing, and compares the existing noisy learning methods, including ELR, Mixup, Dividemix, Unicon, etc. The results show that the method proposed in the present application can still maintain a high recognition accuracy for the power usage behavior pattern under a noise rate of 50%, and has better robustness and generalization performance compared with other methods. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a structure block diagram of the power usage behavior pattern recognition device of the present application;

[0034] Figure 2 is a flowchart of the power usage behavior pattern recognition method of the present application;

[0035] Figure 3 is a contrastive learning diagram of the present application. DETAILED DESCRIPTION

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] like Figure 1 As shown, this embodiment of the invention provides a device for recognizing electricity usage behavior patterns based on noisy learning. It mainly comprises three parts: a module for collecting and preprocessing daily electricity consumption records on the left, a data filtering module based on a Gaussian mixture model in the middle, and a model training module and a pattern recognition module on the right. Specifically, the data collection module collects daily household electricity consumption records for training and testing. The data preprocessing module preprocesses the electricity consumption record data. After preprocessing, the data is fed into two neural networks with identical structures. Preheating training is performed. The data filtering module is used to divide the processed electricity consumption records into a labeled set L and an unlabeled set U using a Gaussian Mixture Model (GMM). The divided data is then input into the neural network, and the model training module calculates the total loss according to the loss function formula. The network parameters are updated via backpropagation. The pattern recognition module uses the trained neural network to identify electricity usage behavior patterns and evaluates the recognition performance.

[0038] Reference Figure 2 The present invention provides a method for recognizing electricity usage behavior patterns based on noisy learning, comprising the following steps:

[0039] (1) Data collection steps: Collect daily electricity consumption records of households, divide the data into training set and test set, and label the training set data according to the usage behavior pattern;

[0040] (2) Data preprocessing steps: preprocess the training set data, including removing outliers, filling missing values, normalization, and data augmentation.

[0041] (3) Model warm-up step: First, the two neural networks are warmed up. The daily electricity consumption records are input into the initial neural network without distinction for preliminary training, so that the neural network has a preliminary ability to recognize electricity usage behavior patterns.

[0042] (4) Data filtering step: The quality of the processed electricity consumption record data is evaluated. The data is divided into two parts by the filtering module, with clean labels and noisy labels respectively, forming a labeled set and an unlabeled set.

[0043] (5) Model training step, using the divided daily electricity record data for model training, calculating the corrected cross-entropy loss using labeled data, calculating the consistency regularization loss using unlabeled data, and using the contrast learning technology to enhance the recognition ability of the electricity record data with unclear behavior pattern characteristics;

[0044] (6) Pattern recognition step, using the trained model to analyze the electricity record data to obtain the recognition result of the power usage behavior pattern.

[0045] In the embodiment of the present application, in the data collection step, the collected data is daily electricity record data of the family, containing readings of 251 families, taking samples at two-minute intervals every month, and reading 24-hour electricity readings. The specific format of the data is: the time step is 46, the dimension is 1, and there are 24 power usage behavior patterns. Among them, 7200 samples are selected for training.

[0046] Data preprocessing refers to preprocessing the input time series, including removing outliers, filling missing values, normalizing, and data enhancement. Among them, the data enhancement includes two means of weak enhancement and strong enhancement, and the concept comes from the paper Fixmatch (see "Sohn K, Berthelot D, Carlini N, et al. Fixmatch: Simplifying semi-supervised learning with consistency and confidence [J]. Advances in neural information processing systems, 2020, 33: 596-608"). Among them, weak enhancement uses MagnitudeWarp: based on a smooth random curve to distort the time series; strong enhancement refers to randomly selecting N (here, 1) from several enhancement strategies (Magnitude Warp, Magnitude Scale, CutOut, Time Noise, Magnitude Noise, Time Warp, Random Crop), then selecting a random amplitude M, and applying the selected enhancement method to the time series data in turn to obtain the transformed time series data.

[0047] In the present application, two neural networks are established for cross-training and updating, and in specific embodiments, common convolutional neural networks or fully connected neural networks can be used. The present application does not limit this.

[0048] In the model preheating step, all mixed labeled daily electricity record data are input into the two neural networks without distinction to preheat the neural networks, and the initial neural network The two neural networks have preliminary recognition ability for power usage behavior patterns.

[0049] In the data screening step, the characteristics of clean data are fitted based on the neural network. According to the loss distribution obtained by the neural network from the daily power consumption record data, the daily power consumption record data is divided into two parts with clean labels and with noisy labels. In the model training step, in order to avoid the neural network from generating high confidence for the noise time series, the co-teaching training strategy is used, that is, the two neural networks are trained by the clean data and the noisy data respectively. The screened time series set D (1) ={L (1) ,U (1)} is used for training the neural network in the next stage, and the neural network The screened time series set D (2) ={L (2) ,U (2)} is used for training the neural network in the next stage. The specific process of data screening and model training is shown in Figure 3 , including:

[0050] i. The power consumption record data is input into the two neural networks to obtain the cross-entropy loss of each, and the loss distribution is modeled using a Gaussian mixture model, that is, a linear combination of K Gaussian distribution functions, each part is represented by gi, here K is 2, that is, corresponding to g1, g2, the two parts g1, g2 correspond to power consumption record data with clean labels and with noisy labels respectively, and the parameters of the Gaussian mixture model are solved by the EM algorithm.

[0051] For the normalized loss l obtained by the neural network for clean samples and noise samples, l ∈ [0, 1], assuming that its distribution is composed of two Gaussian distributions, then the corresponding probability density function (pdf) can be defined as:

[0052]

[0053] Where ω k is the coefficient corresponding to the probability density function of each component, ω k = 1, 0 ≤ ω k ≤ 1, μ k is the expected vector corresponding to each component, Σ k is the covariance matrix corresponding to each component, ω k , μ k , Σ k are the parameters to be solved, which can be solved by the EM algorithm.

[0054] Next, the probability that the loss corresponding to each electricity consumption record belongs to different parts of the Gaussian model is calculated. Based on the set filtering threshold, the electricity consumption record data is divided into clean data and noisy data, which are respectively designated as the labeled set L and the unlabeled set U, forming the dataset D = {L, U}. Two neural networks The two datasets obtained after filtering are denoted as D. (1) ={L (1) U (1)}, D (2) ={L (2) ,u (2) Because in each round of training, the original training data is input into two separate neural networks. Therefore, during the filtering stage, two partitioned datasets will be obtained, namely d. (1) ={l (1) U (1)}, d (2) ={l (2) U (2) L represents the labeled set consisting of samples considered clean, and U represents the unlabeled set. Note the D here. (1) D (2) This refers only to the data after it has been re-divided, without any data augmentation. Data augmentation is required before training on the re-divided data. Figure 3 Enhanced data copy 1 and enhanced data copy 2 refer to the data obtained after the time series data has undergone weak enhancement and strong enhancement transformation, respectively.

[0055] ii. Cross-training: The Adjmatch-based method is used to train the noisy electricity consumption records. The dataset D... (1) D (2) The inputs are cross-referenced and fed into another neural network for training, and the loss is calculated according to the formula. Including the corrected supervised cross-entropy loss Consistency regularization loss and comparative loss Simultaneously, optimizers such as SGD are used to update parameters.

[0056] Specifically, in the training stage, in order to avoid the proportion imbalance of the behavior pattern category corresponding to the screened daily power consumption record data with clean labels, which leads to the deviation of the recognition ability of the neural network, the present application introduces the category prior information to correct the supervised loss. In order to avoid too little daily power consumption record data used for training, for the remaining daily power consumption record data regarded as having noisy labels, the present application adds the training by a consistency regular unsupervised constraint, controls the quality of the data used by a threshold, and thus fully utilizes the data. Considering that there is still data in the daily power consumption record that is difficult to learn the power use behavior pattern feature, in order to increase the robustness of the neural network, the present application introduces contrastive learning, learns the invariance of the power use behavior pattern after data enhancement, and improves the robustness of the neural network to the power use behavior pattern recognition.

[0057] According to the embodiment of the present application, the formula of the overall loss function is:

[0058]

[0059] where λ u and λ c are hyperparameters, representing the weights of the unsupervised loss and the contrastive loss, respectively.

[0060] The specific formula of the cross-entropy loss L corrected by the category is:

[0061]

[0062] where L is the set of labeled data, π y is the estimate of the prior , τ is a hyperparameter, τ > 0. f y (x) represents the logit output of the sample x predicted by the neural network as the label y corresponds, and ∑ y′ represents the summation of all label categories.

[0063] The specific formula of the consistency regularization loss L is:

[0064]

[0065] where H(p||q) represents the cross-entropy loss of two distributions, α(x) represents a weak enhancement method, represents a strong enhancement method, q θ (·) represents the prediction result of the network output, and T represents the temperature coefficient.

[0066] Specifically, for a processed power consumption record data x i, respectively, after weak and strong augmentation, the transformed data augmentation copy α(x i ), The data augmentation copy α(x i ) obtained by weak augmentation is sequentially passed through the encoder and the classifier to obtain the predicted class distribution q b of the neural network. Then, argmax(q b ) is taken as the pseudo label Finally, it is used to calculate the cross-entropy loss of the data augmentation copy obtained by strong augmentation.

[0067] Where the unsupervised contrastive loss Lunsupervisedis calculated includes two parts, namely, the spatiotemporal contrastive loss and the instance contrastive loss. The spatiotemporal contrastive loss is used to learn the consistency of the daily power consumption record data in space and time, and the instance contrastive loss is used to learn the individual consistency of the daily power consumption record data.

[0068] The specific formula of the spatiotemporal contrastive loss Lspatiotemporalis as follows:

[0069]

[0070]

[0071] Where i represents the index of the time series data, t represents the time step, N represents the number of data in a batch, represents the sum of all time steps of a time series data, z i,t represents the feature vector segment obtained by nonlinear transformation of the i-th data at the t-th time step after weak augmentation, z′ i,t represents the feature vector segment obtained by nonlinear transformation of the i-th data at the t-th time step after strong augmentation, sim(z i,t ,z′ i,t ) represents the similarity between z i,t and z′ i,t , which can be calculated by cosine similarity.

[0072] The specific formula of the instance contrastive loss Linstanceis as follows:

[0073]

[0074]

[0075] represents the sum of B data in a batch.

[0076] ​​Overall unsupervised contrastive loss The specific formula is:

[0077]

[0078] Wherein, alpha is a hyperparameter, indicating l inst And l temp The proportion of the proportion.

[0079] In the mode recognition step, when the power usage behavior mode is recognized, the two neural networks The comprehensive recognition result, for example, taking the average value of the two neural network prediction results as the final recognition result of the power usage behavior mode, improves the stability of the recognition effect. Additionally, the recognition ability of the power usage behavior mode can be measured by accuracy, recall rate, precision, etc.

[0080] In actual implementation, common deep learning frameworks such as Tensorflow, Pytorch, etc. can be used to implement the method of the present application.

[0081] The present application collects actual daily household electricity consumption record data for training and testing, and compares existing noisy learning methods, including ELR, Mixup, Dividemix, Unicon, etc. The test accuracy and wAUC of each method are shown in Table 1, and the results show that the method proposed in the present application can still maintain a high recognition accuracy for the power usage behavior mode when the noise rate is as high as 50%, and has better robustness and generalization performance compared with other methods.

[0082] Table 1 Test accuracy and wAUC of each method when noise rate is 50%

[0083] Method Accuracy wAUC Adjmatch (the invention) 0.6883 0.9149 ELR 0.6721 0.9143 MixUp 0.6336 0.9124 Dividemix 0.6116 0.8737 Unicon 0.5746 0.8680

[0084] The wAUC described in the present application is weight AUC, also known as weighted AUC.

[0085] The present application also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs are executed by the processor to realize the steps of the power usage behavior mode recognition method based on noisy learning as described above.

[0086] The present application also provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by the processor to realize the steps of the power usage behavior mode recognition method based on noisy learning as described above.

[0087] The application provides a power usage behavior pattern recognition method and device based on noisy learning, which can recognize the power usage behavior pattern through processing and analyzing the daily power consumption record data of a family, can overcome the data noise problem, and has high accuracy and robustness. The application can help the family better manage the power usage, improve the energy utilization efficiency, and has wide application prospect.

Claims

1. A power usage behavior pattern recognition method based on noisy learning, characterized by, The method comprises the following steps: a data collection step of collecting daily power consumption record data of a family, labeling the data according to the use behavior mode thereof, and forming a training data set; a data preprocessing step of preprocessing the training set data, including removing outliers, filling missing values, normalizing, and data enhancement, wherein the data enhancement includes two modes of weak enhancement and strong enhancement; a model preheating step of preliminarily training daily power consumption records in the training data set in two neural networks without distinction, so that the neural network has a preliminary recognition ability for the power use behavior mode; a data screening step of screening the power consumption record data using a Gaussian mixture model, dividing the data into two parts, respectively with clean labels and with noise labels, and forming a labeled set and an unlabeled set; a model training step of using the divided daily power consumption record data for model training, calculating a corrected cross-entropy loss using the labeled data, calculating a consistency regularization loss using the unlabeled data, and obtaining refined behavior mode features by contrastive learning; In the training step, the co-teaching training method is adopted, two neural networks are used for collaborative training, the time series screened by one neural network will be handed over to another neural network for training, wherein the class corrected supervised cross entropy loss is used for the labeled data, the consistency regularization loss is used for the unlabeled data, and the contrast learning is used to strengthen the transformation consistency of the power consumption record data, and the total loss in the training is The supervised cross entropy loss including correction is obtained The consistency regularization loss And the contrast loss At the same time, the SGD optimizer is used for parameter updating; the form of the contrast loss function is as follows: wherein a is a hyper-parameter, the spatio-temporal contrastive loss The specific formula is: Example contrastive loss The specific formula is: where i denotes the index of time series data, t denotes the time step, N denotes the number of data in a batch, denotes the sum of all time steps of a time series data, z i,t denotes the feature vector segment of the i-th data at the t-th time step after weak enhancement and nonlinear transformation by neural network, z ′ i,t denotes the feature vector segment of the i-th data at the t-th time step after strong enhancement and nonlinear transformation by neural network, sim(z i,t ,z ′ i,t denotes the similarity between z i,t and z ′ i,t , τ is a hyperparameter; a mode recognition step of analyzing the power consumption record data using the trained model to obtain a recognition result of the power use behavior mode.

2. The power usage behavior pattern recognition method based on noisy learning according to claim 1, wherein, In the data screening step, a two-part Gaussian mixture model is used to model the loss distribution of the daily power consumption record data obtained by the neural network, the Gaussian mixture model parameters are solved by the EM algorithm, then the probability of each power consumption record data belonging to each Gaussian part is calculated, and the clean data is screened out as the labeled data according to the set screening threshold, and the remaining power consumption record data is used as the unlabeled data.

3. The power usage behavior pattern recognition method based on noisy learning according to claim 2, wherein, The two-part Gaussian mixture model modeling includes: Normalized loss obtained by neural network for clean sample and noise sample The distribution is composed of two Gaussian distributions, and the corresponding probability density function is defined as: where ω k is the coefficient corresponding to the probability density function of each component, 0≤ω k ≤1, μ k is the expectation vector corresponding to each component, Σ k is the covariance matrix corresponding to each component, ω k , μ k , Σ k is the parameter to be solved.

4. The power usage behavior pattern recognition method based on noisy learning of claim 1, wherein, a supervised cross-entropy loss with category correction: where L is a set of labeled data, π y is an estimate of the prior , τ is a hyperparameter, τ > 0, f y (x) represents the logit output predicted by the neural network for sample x to correspond to label y, and ∑ y′ represents a summation over all label classes.

5. The noisy-label based power usage behavior pattern recognition method of claim 1, wherein, a consistency regularization loss function form is as follows: wherein U represents a set of unlabeled data, H(p||q) represents a cross-entropy loss of two distributions, a(x) represents a data augmentation copy obtained by a weak augmentation manner, represents a data augmentation copy obtained by a strong augmentation manner, q θ (·) represents a prediction result output by a neural network, and T represents a temperature coefficient.

6. A power usage behavior pattern recognition device based on noisy learning, characterized by, including: a data collection module for collecting daily power consumption record data of a family, labeling the data according to the use behavior mode thereof, and forming a training data set; a data preprocessing module for preprocessing the training set data, including removing outliers, filling missing values, normalizing, and data enhancement, wherein the data enhancement includes two modes of weak enhancement and strong enhancement; a model preheating module for preliminarily training daily power consumption records in the training data set in two neural networks without distinction, so that the neural network has a preliminary recognition ability for the power use behavior mode; a data screening module for screening the power consumption record data using a Gaussian mixture model, dividing the data into two parts, respectively with clean labels and with noise labels, and forming a labeled set and an unlabeled set; The model training module is configured to use the divided daily power consumption record data for model training, calculate a corrected cross-entropy loss using labeled data, calculate a consistency regularization loss using unlabeled data, and obtain refined behavior pattern features using contrastive learning. The model training module uses a co-teaching training method to use two neural networks for collaborative training. A neural network selected time series is handed over to another neural network for training. For labeled data, a class-corrected supervised cross-entropy loss is used. For unlabeled data, a consistency regularization loss is used. Contrastive learning is used to strengthen the consistency of the transformation of the power consumption record data. In the training, the total loss includes a corrected supervised cross-entropy loss a consistency regularization loss and a contrastive loss The parameters are updated using an SGD optimizer. The contrastive loss function is as follows: wherein a is a hyper-parameter, the spatio-temporal contrastive loss The specific formula is: Example contrastive loss The specific formula is: where i denotes the index of time series data, t denotes the time step, N denotes the number of data in a batch, denotes the sum of all time steps of a time series data, z i,t denotes the feature vector segment of the i-th data at the t-th time step after weak enhancement processing and nonlinear transformation by the neural network, z ′ i,t denotes the feature vector segment of the i-th data at the t-th time step after strong enhancement processing and nonlinear transformation by the neural network, sim(z i,t ,z ′ i,t denotes the similarity between z i,t and z ′ i,t , τ is a hyperparameter; a mode recognition module for analyzing the power consumption record data using the trained model to obtain a recognition result of the power use behavior mode.

7. A computer device, characterized by including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs are executed by the processor to implement the steps of the power use behavior mode recognition method based on noisy learning according to any one of claims 1-5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, which is executed by a processor, implements the steps of the power usage behavior pattern recognition method based on noisy learning as claimed in any one of claims 1-5.

Images