Non-intrusive load identification method based on comparative learning

Abstract

The invention discloses a non-intrusive load identification method based on comparative learning. The non-intrusive load identification method comprises the following steps: loading an original training data set and initializing parameters; constructing a multi-view training data set; dividing a multi-view training data set; reading a first batch of training data, inputting the first batch of training data into an encoder, and extracting high-dimensional features of data samples; obtaining supervised classification loss; unsupervised loss is obtained; obtaining image comparison loss; calculating the gradient of the comprehensive loss relative to the model parameters, and updating the model parameters; and repeating the steps for multiple rounds until a preset number of iterations or convergence precision is met, and finally obtaining a load identification model. According to the method, self-supervised training can be performed on the load identification model by using the label-free load data samples, so that the dependence on a large number of label data samples is effectively reduced; the introduced graph comparison learning framework improves the capability of learning identification features of the model by maximizing the similarity between the pseudo-tag graph and the embedded feature graph, thereby improving the load identification precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of non-intrusive power load monitoring, and in particular relates to a non-intrusive load identification method based on contrast learning. Background Art

[0002] Under the current "dual carbon" background, the demand side of electricity is maintaining a rigid upward trend. A large number of studies have shown that if the flexibility of the power system is extended to the user side, that is, the end users or power consumers in the power system, so that users can more actively participate in power dispatching and load adjustment by adjusting their own power consumption behavior, it can effectively support the balanced and stable operation of the power system and promote the integration of renewable energy. The key to achieving this task is to provide users with more fine-grained power consumption information, so as to ensure that users can flexibly and effectively adjust their power consumption behavior to improve energy efficiency. On this basis, the development of accurate load monitoring technology has become a research direction of common concern in various countries, such as Liu Bo, Luan Wenpeng. Power consumption data cloud architecture solution and application scenarios based on load decomposition [J]. Power System Technology, 2016, 40(03): 791-796.

[0003] Non-intrusive Load Monitoring (NILM) technology collects the total power consumption data at the user's entrance bus, and can obtain the energy consumption information and operating status of each power device through algorithm decomposition, providing a simple and cost-effective method for load monitoring. NILM technology enables smart grids to make more accurate predictions of energy demand, thereby providing a more stable power supply for all users of the power grid, such as Yu Yixin, Liu Bo, Luan Wenpeng. Non-intrusive residential power load monitoring and decomposition technology [J]. China Southern Power Grid Technology, 2013, 7(04): 1-5.

[0004] In recent years, with the rapid development of computer technology and artificial intelligence big data, deep learning, as a deeper machine learning model, has been favored by researchers for its powerful feature learning and data fitting capabilities. Now its application field has gradually expanded to NILM technology and has flourished. Given that most deep learning models focus on supervised learning and require a large amount of labeled data support, and as a branch of deep learning, the concept of contrastive learning has spawned many studies that can learn image-level representations without relying on manual annotations, such as Kelly J and Knottenbelt W, "Neural NILM: Deep Neural Networks Applied to Energy Disaggregation," in Pro. 2nd ACM Int. Conf. on Embed. Syst. for Energy-Effi. Built Environ., New York, NY, USA, 2015, pp. 55-64. Therefore, in order to reduce the dependence of the NILM method based on deep neural networks on a large number of labeled data samples, the present invention adopts a contrastive learning method based on graph regularization to perform self-supervised training on the load identification model based on unlabeled load data samples, in order to improve the feature learning ability of the model, thereby improving the load identification accuracy. Summary of the invention

[0005] The present invention is proposed to solve the problems existing in the prior art, and its purpose is to provide a non-intrusive load identification method based on contrast learning.

[0006] The technical solution of the present invention is: a non-intrusive load identification method based on contrastive learning, comprising the following steps:

[0007] A. Load the original training data set and initialize the parameters;

[0008] B. Construct a multi-view training dataset;

[0009] C. Divide the multi-view training dataset;

[0010] D. Read the first batch of training data, input it into the encoder, and extract the high-dimensional features of the data samples;

[0011] E. For labeled training data samples, calculate the cross entropy and obtain the supervised classification loss;

[0012] F. For unlabeled training data samples, calculate the cross entropy and obtain the unsupervised loss;

[0013] G. Get the image comparison loss;

[0014] H. Calculate the gradient of the comprehensive loss relative to the model parameters and update the model parameters;

[0015] I. Determine whether the current batch training has reached the preset number of iterations or meets the preset convergence accuracy, end the training and output or return to step E for the next batch training.

[0016] Furthermore, step A loads the original training data set and initializes the parameters. The specific process is as follows:

[0017] First, during the loading process of the original training dataset, a training dataset containing a small number of labeled and a large number of unlabeled voltage-current trajectory image samples is loaded;

[0018] Then, the parameters of the encoder, classifier and mapper included in the load identification model are initialized;

[0019] Finally, the parameters in the optimizer used for load identification model training are initialized.

[0020] Furthermore, step B constructs a multi-view training dataset. The specific process is as follows:

[0021] First, a random augmentation transformation is performed on each data sample in the loaded training dataset;

[0022] Then, a weak enhancement transformation is performed on the labeled voltage-current trajectory image samples;

[0023] Then, the unlabeled voltage-current trajectory image samples are processed by weak enhancement transformation and two strong enhancement transformations respectively;

[0024] Finally, the enhanced transformation is used to derive new views for the voltage-current trajectory image samples, and multiple views of all the voltage-current trajectory image samples together constitute a new multi-view training dataset. The label information of the voltage-current trajectory image samples is directly assigned to their derived new view samples.

[0025] Furthermore, step C divides the multi-view training data set, and the specific process is as follows:

[0026] The multi-view training data set constructed in step B is randomly divided into several small batch training data sets, and the small batch training data sets are used for subsequent load identification model training.

[0027] Furthermore, step E calculates the cross entropy for the labeled training data samples to obtain the supervised classification loss. The specific process is as follows:

[0028] First, for the labeled training data samples in the mini-batch training data set, the classifier is used to generate its class probability;

[0029] Then, the cross entropy between the class probability and the true label information is calculated to obtain the supervised classification loss.

[0030] Furthermore, step F calculates the cross entropy for the unlabeled training data samples to obtain the unsupervised loss. The specific process is as follows:

[0031] First, for the unlabeled training data samples in the mini-batch training dataset, the classifier and mapper are used to generate their pseudo labels;

[0032] Then, the cross entropy between the pseudo labels and their class probabilities is calculated to obtain the unsupervised loss.

[0033] Furthermore, step G obtains the image contrast loss, and the specific process is as follows:

[0034] First, construct the pseudo label map and embedded feature map of the current batch of training samples;

[0035] Then, the cross entropy between the pseudo label map and the embedded feature map is calculated;

[0036] Finally, the image contrast loss is obtained.

[0037] Furthermore, step H calculates the gradient of the comprehensive loss relative to the model parameters and updates the model parameters. The specific process is as follows:

[0038] First, the weighted sum of supervised classification loss, unsupervised loss, and image contrast loss is calculated to obtain the comprehensive loss function;

[0039] Then, using the back-propagation algorithm, the gradient of the comprehensive loss with respect to the model parameters is calculated;

[0040] Finally, according to the obtained gradient, the stochastic gradient descent method is used to update the model parameters.

[0041] Furthermore, step I determines whether the current batch training reaches a preset number of iterations or meets a preset convergence accuracy, ends the training and outputs or returns to step E for the next batch training. The specific process is as follows:

[0042] First, determine whether the current batch training has reached the preset number of iterations or meets the preset convergence accuracy;

[0043] Then, if the preset conditions are met, the encoder, classifier and mapper obtained by the current training are output to obtain the final load identification model, and the training process ends;

[0044] Then, if the preset conditions are not met, the training data in the next batch of small batch training data sets is read and go to step E.

[0045] The beneficial effects of the present invention are as follows:

[0046] In order to solve the problem that the traditional load identification model training process based on semi-supervised learning is sensitive to noise samples in unlabeled data sets and easily leads to the accumulation of prediction bias, the present invention designs a non-intrusive load identification method based on contrastive learning, which can use unlabeled load data samples to perform self-supervised training on the load identification model, effectively reducing the dependence on a large number of labeled data samples; the introduced graph contrastive learning framework improves the model's ability to learn identification features by maximizing the similarity between the pseudo-label graph and the embedded feature graph, thereby improving the load identification accuracy of the NILM model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of a non-intrusive load identification method based on contrastive learning of the present invention;

[0048] Figure 2 It is a training process framework of a load identification model based on contrastive learning designed by the present invention. DETAILED DESCRIPTION

[0049] Hereinafter, the present invention will be described in detail with reference to the accompanying drawings and embodiments:

[0050] like Figure 1 to Figure 2 As shown, the non-intrusive load identification method based on contrastive learning includes the following steps:

[0051] A. Load the original training data set and initialize the parameters;

[0052] B. Construct a multi-view training dataset;

[0053] C. Divide the multi-view training dataset;

[0054] D. Read the first batch of training data, input it into the encoder, and extract the high-dimensional features of the data samples;

[0055] E. For labeled training data samples, calculate the cross entropy and obtain the supervised classification loss;

[0056] F. For unlabeled training data samples, calculate the cross entropy and obtain the unsupervised loss;

[0057] G. Get the image comparison loss;

[0058] H. Calculate the gradient of the comprehensive loss relative to the model parameters and update the model parameters;

[0059] I. Determine whether the current batch training has reached the preset number of iterations or meets the preset convergence accuracy, end the training and output or return to step E for the next batch training.

[0060] Step A loads the original training data set and initializes the parameters. The specific process is as follows:

[0061] First, during the loading process of the original training dataset, a training dataset containing a small number of labeled and a large number of unlabeled voltage-current trajectory image samples is loaded;

[0062] Then, the parameters of the encoder, classifier and mapper included in the load identification model are initialized;

[0063] Finally, the parameters in the optimizer used for load identification model training are initialized.

[0064] Step B constructs a multi-view training dataset. The specific process is as follows:

[0065] First, a random augmentation transformation is performed on each data sample in the loaded training dataset;

[0066] Then, a weak enhancement transformation is performed on the labeled voltage-current trajectory image samples;

[0067] Then, the unlabeled voltage-current trajectory image samples are processed by weak enhancement transformation and two strong enhancement transformations respectively;

[0068] Finally, the enhanced transformation is used to derive new views for the voltage-current trajectory image samples, and multiple views of all the voltage-current trajectory image samples together constitute a new multi-view training dataset. The label information of the voltage-current trajectory image samples is directly assigned to their derived new view samples.

[0069] Step C divides the multi-view training data set. The specific process is as follows:

[0070] The multi-view training data set constructed in step B is randomly divided into several small batch training data sets, and the small batch training data sets are used for subsequent load identification model training.

[0071] Step E: For labeled training data samples, calculate the cross entropy and obtain the supervised classification loss. The specific process is as follows:

[0072] First, for the labeled training data samples in the mini-batch training data set, the classifier is used to generate its class probability;

[0073] Then, the cross entropy between the class probability and the true label information is calculated to obtain the supervised classification loss.

[0074] Step F calculates the cross entropy for the unlabeled training data sample and obtains the unsupervised loss. The specific process is as follows:

[0075] First, for the unlabeled training data samples in the mini-batch training dataset, the classifier and mapper are used to generate their pseudo labels;

[0076] Then, the cross entropy between the pseudo labels and their class probabilities is calculated to obtain the unsupervised loss.

[0077] Step G obtains the image contrast loss, and the specific process is as follows:

[0078] First, construct the pseudo label map and embedded feature map of the current batch of training samples;

[0079] Then, the cross entropy between the pseudo label map and the embedded feature map is calculated;

[0080] Finally, the image contrast loss is obtained.

[0081] Step H calculates the gradient of the comprehensive loss relative to the model parameters and updates the model parameters. The specific process is as follows:

[0082] First, the weighted sum of supervised classification loss, unsupervised loss, and image contrast loss is calculated to obtain the comprehensive loss function;

[0083] Then, using the back-propagation algorithm, the gradient of the comprehensive loss with respect to the model parameters is calculated;

[0084] Finally, according to the obtained gradient, the stochastic gradient descent method is used to update the model parameters.

[0085] Step I determines whether the current batch training reaches the preset number of iterations or meets the preset convergence accuracy, ends the training and outputs or returns to step E for the next batch training. The specific process is as follows:

[0086] First, determine whether the current batch training has reached the preset number of iterations or meets the preset convergence accuracy;

[0087] Then, if the preset conditions are met, the encoder, classifier and mapper obtained by the current training are output to obtain the final load identification model, and the training process ends;

[0088] Then, if the preset conditions are not met, the training data in the next batch of small batch training data sets is read and go to step E.

[0089] Specifically, step A loads the original training data set and initializes the parameters, as follows:

[0090] First, the voltage-current trajectory image samples are drawn from the voltage and current waveform sampling data during the stable operation period before and after the electrical equipment is turned on.

[0091] Then, the collected raw data of voltage and current waveforms are preprocessed by using the following equations (1) and (2), and then a voltage-current trajectory image is drawn with V and I as the vertical and horizontal axes respectively.

[0092] V=(V on +V off ) / 2(1)

[0093] I=I on -I off (2)

[0094] In formula (1), V on and I on are the actual measured values ​​of voltage and current before the electrical equipment is turned on, V off and I off It is the actual measured value of the voltage after turning on.

[0095] Specifically, the weak enhancement transformation in step B is any one of random cropping and random rotation, and the strong enhancement transformation refers to grayscale transformation.

[0096] The two strong enhancement transformations in the present invention select different image pixel grayscale values.

[0097] Specifically, in step D, a convolutional neural network is selected as the encoder, and represented by f(x), the parameters of the encoder are updated during the iterative training of the input multi-view training dataset.

[0098] Specifically, in step E, the classifier is composed of a fully connected layer and a Softmax function, and the mapper is composed of a multi-layer perceptron (MLP). The two jointly receive the high-dimensional features output by the convolutional neural network to obtain the class probability distribution and embedded feature vector of the multi-view training samples, respectively. The two are calculated by equations (3) and (4) respectively.

[0099] p(y|x)=h·f(x)(3)

[0100] z=g·(f(x))(4)

[0101]

[0102] In the formula, the classifier is represented by h(·) and the mapper is represented by g(·). The supervised classification loss function is obtained by formula (5), and B represents the total number of samples in the current training batch.

[0103] Specifically, in step F, the class probabilities generated by the classifier and the embedded features generated by the mapper are used to obtain pseudo labels for unlabeled training data samples by optimizing a specific function.

[0104] The first term of the above function is a smooth constraint based on the embedded feature, which aims to make the pseudo label of the current sample as close as possible to the class probability of other similar samples in the same batch. The definition of similarity quantifies the relative similarity between the current sample and each sample in the batch by calculating the inner product of the two corresponding embedded feature vectors in the embedding space.

[0105] The second term of the above function is to keep the pseudo label of the current sample as close to its original class probability as possible. The whole process takes into account global information and local characteristics, meeting the requirements of the smoothness constraint (as close as possible to the class probability of similar samples) without deviating too far from the original class probability.

[0106] The above function is specifically shown in formula (6), using a k The similarity between the current sample and the kth sample in the training batch is calculated by formula (7). k has been normalized, the optimization process shown in formula (6) can be further derived into a direct formula for pseudo-labels as shown in formula (8).

[0107] The unsupervised classification loss function is specifically obtained through formula (8).

[0108]

[0109] In equation (6), t is a scalar temperature parameter. In equation (7), α is used to adjust the influence of the two terms on the right side of the equation on the objects on the left side. In equation (8), μ defines the relative size of labeled and unlabeled samples in the current batch of training samples, and discards pseudo labels whose class probability distribution is less than the threshold τ.

[0110] Specifically, in step G, a pseudo-label map and an embedded feature map are established by constructing a similarity matrix, with the goal of minimizing the map contrast loss. The mapper is trained under the guidance of the generated pseudo-label map so that the data points of samples (positive sample pairs) with similar pseudo-labels are close to each other in the embedding space, thereby outputting similar embedded features, thereby enhancing the feature extraction capability of the mapper; these similar embedded features can further implement smoothness constraints on the classifier to improve the quality of the pseudo-labels.

[0111] The image contrast loss function is obtained by formula (9).

[0112]

[0113] In the formula, the two graphs represent and Normalize.

[0114] Specifically, the comprehensive loss function in step H is expressed by formula (10), and the parameters of the encoder, classifier and mapper in the load identification model are trained by minimizing the comprehensive loss value.

[0115] L=L x +ηL u1 +βL u2 (10)

[0116] Where η and β are scalar hyperparameters, which are used to control the weights of unsupervised classification loss and graph contrast loss respectively.

[0117] Specifically, in step I, the iterative training rounds may be set to 70, and the learning rate may be set to 0.0008, but is not limited thereto. Stochastic gradient descent (SGD) with a momentum parameter of 0.9 and a weight decay of 0.0001 is selected for training, and the learning rate is adjusted by a cosine annealing decay strategy.

[0118] Specifically, the present invention extracts a large number of labeled and a small number of unlabeled voltage-current trajectory images based on the current and voltage data of the user bus end, and uses them as the input of the NILM model to construct a non-intrusive load identification model based on contrastive learning, and improves the identification accuracy of the NILM method through contrastive learning between pseudo-label images and embedded feature maps.

[0119] Specifically, the method of the present invention comprises a data preprocessing module, a pseudo label training module and a model output module. The first is the data preprocessing module, then the pseudo label training module, and finally the model output module. The detailed flowchart is as follows: Figure 1 As shown, the pseudo-label training module aims to train the mapper to learn features that better characterize the user's electrical equipment category, and use these similar features to impose smooth constraints on the classifier to improve the labeling accuracy. First, two device representations, class probability distribution and low-dimensional embedded feature vectors, are obtained based on the high-dimensional features of each electrical equipment sample; then, the pseudo-labels of unlabeled electrical equipment are calculated based on the two representations; the established graph contrast loss aims to extract the key structure and semantic information of the pseudo-label graph and the embedded feature graph. Specifically, the embedded features are regularized by pseudo-labels to continuously improve the feature learning ability of the model. At the same time, the pseudo-labels are improved by smooth constraints of the embedded features, which significantly reduces the prediction bias, thereby achieving good performance in multi-class electrical equipment recognition tasks.

[0120] Verification Example

[0121] Taking the actual application scenario as an example, the specific steps of implementing semi-supervised non-intrusive load identification using the method of the present invention are as follows:

[0122] First, all samples in the dataset are randomly divided into five subsets of equal size, and the five-fold cross-validation method is used. In each experiment, four subsets are used as training sets, and the remaining subset is used as a test set. Finally, the results of the five tests are summarized for comprehensive evaluation.

[0123] Taking 8 residential users in Pittsburgh, Pennsylvania, USA as an example, 8 measuring devices are deployed to collect the total user and internal electrical appliance power data, and all data are divided into training and test data sets in a ratio of 4:1. 11 target electrical appliances are selected, namely: air conditioners, compact fluorescent lamps, electric fans, refrigerators, hair dryers, heaters, incandescent bulbs, laptops, microwave ovens, vacuum cleaners, washing machines, etc. Figure 2 Load identification is performed under the semi-supervised load identification neural network structure based on graph contrast regularization as shown.

[0124] Then, the parameters of the semi-supervised load identification neural network are initialized, and the normalized voltage-current trajectory image is input into the network structure for training to obtain the network model.

[0125] Finally, the trained model is tested, and the voltage-current trajectory image converted from the total power data of the test data set is input into the trained model to perform load identification on the target electrical equipment in the test set.

[0126] In this research material example, the evaluation index F for the load identification result is score , whose expression is as follows:

[0127]

[0128]

[0129] The MAE and SAE results of load decomposition using the method of the present invention with labeled data are shown in Table 1.

[0130] Table 111 F of identification results of electrical equipment score index

[0131]

[0132] It can be seen from the above research materials that the method of the present invention can accurately identify the device category.

[0133] Although the present invention has been described above in conjunction with the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can make many modifications without departing from the purpose of the present invention, which are all within the protection of the present invention.

Claims

1. A non-intrusive load identification method based on contrastive learning, characterized by: The following steps are involved: A. Load the original training data set and initialize the parameters; B. Construct a multi-view training dataset; C. Divide the multi-view training dataset; D. Read the first batch of training data, input it into the encoder, and extract the high-dimensional features of the data samples; E. For labeled training data samples, calculate the cross entropy and obtain the supervised classification loss; F. For unlabeled training data samples, calculate the cross entropy and obtain the unsupervised loss; G. Get the image comparison loss; H. Calculate the gradient of the comprehensive loss relative to the model parameters and update the model parameters; I. Determine whether the current batch training has reached the preset number of iterations or meets the preset convergence accuracy, end the training and output or return to step E for the next batch training.

2. The non-intrusive load identification method based on contrastive learning according to claim 1 is characterized in that: Step A loads the original training data set and initializes the parameters. The specific process is as follows: First, during the loading process of the original training dataset, a training dataset containing a small number of labeled and a large number of unlabeled voltage-current trajectory image samples is loaded; Then, the parameters of the encoder, classifier and mapper included in the load identification model are initialized; Finally, the parameters in the optimizer used for load identification model training are initialized.

3. The non-intrusive load identification method based on contrastive learning according to claim 1, characterized in that: Step B constructs a multi-view training dataset. The specific process is as follows: First, a random augmentation transformation is performed on each data sample in the loaded training dataset; Then, a weak enhancement transformation is performed on the labeled voltage-current trajectory image samples; Then, the unlabeled voltage-current trajectory image samples are processed by weak enhancement transformation and two strong enhancement transformations respectively; Finally, the enhanced transformation is used to derive new views for the voltage-current trajectory image samples, and multiple views of all the voltage-current trajectory image samples together constitute a new multi-view training dataset. The label information of the voltage-current trajectory image samples is directly assigned to their derived new view samples.

4. The non-intrusive load identification method based on contrastive learning according to claim 1, characterized in that: Step C divides the multi-view training data set. The specific process is as follows: The multi-view training data set constructed in step B is randomly divided into several small batch training data sets, and the small batch training data sets are used for subsequent load identification model training.

5. The non-intrusive load identification method based on contrastive learning according to claim 1, characterized in that: Step E: For labeled training data samples, calculate the cross entropy and obtain the supervised classification loss. The specific process is as follows: First, for the labeled training data samples in the mini-batch training data set, the classifier is used to generate its class probability; Then, the cross entropy between the class probability and the true label information is calculated to obtain the supervised classification loss.

6. The non-intrusive load identification method based on contrastive learning according to claim 1, characterized in that: Step F calculates the cross entropy for the unlabeled training data sample and obtains the unsupervised loss. The specific process is as follows: First, for the unlabeled training data samples in the mini-batch training dataset, the classifier and mapper are used to generate their pseudo labels; Then, the cross entropy between the pseudo labels and their class probabilities is calculated to obtain the unsupervised loss.

7. The non-intrusive load identification method based on contrastive learning according to claim 1, characterized in that: Step G obtains the image contrast loss, and the specific process is as follows: First, construct the pseudo label map and embedded feature map of the current batch of training samples; Then, the cross entropy between the pseudo label map and the embedded feature map is calculated; Finally, the image contrast loss is obtained.

8. The non-intrusive load identification method based on contrastive learning according to claim 1, characterized in that: Step H calculates the gradient of the comprehensive loss relative to the model parameters and updates the model parameters. The specific process is as follows: First, the weighted sum of supervised classification loss, unsupervised loss, and image contrast loss is calculated to obtain the comprehensive loss function; Then, using the back-propagation algorithm, the gradient of the comprehensive loss with respect to the model parameters is calculated; Finally, according to the obtained gradient, the stochastic gradient descent method is used to update the model parameters.

9. The non-intrusive load identification method based on contrastive learning according to claim 1, characterized in that: Step I determines whether the current batch training reaches the preset number of iterations or meets the preset convergence accuracy, ends the training and outputs or returns to step E for the next batch training. The specific process is as follows: First, determine whether the current batch training has reached the preset number of iterations or meets the preset convergence accuracy; Then, if the preset conditions are met, the encoder, classifier and mapper obtained by the current training are output to obtain the final load identification model, and the training process ends; Finally, if the preset conditions are not met, read the training data in the next batch of small batch training data sets and go to step E.

Images