CN114359674A - A Non-Intrusive Load Identification Method Based on Metric Learning - Google Patents

Abstract

The invention provides a non-invasive load identification method based on metric learning. The method can realize effective identification of unknown load and has strong generalization capability; on the other hand, metric learning, which is one of the methods for small sample learning, can reduce the dependence on training samples, and has high practicability.

Description

A Non-Intrusive Load Identification Method Based on Metric Learning

technical field

The present invention relates to the field of non-intrusive load monitoring (NILM), in particular to a non-intrusive load identification method based on metric learning.

Background technique

With the rapid development of society, the demand for energy continues to increase, and electric energy, as the main secondary energy, is one of the main energy utilization methods. How to improve the efficiency of power use and realize intelligent power consumption has attracted widespread attention, and it is particularly important to master the detailed power consumption information on the user side. The traditional intrusive load monitoring requires the installation of acquisition and communication devices in each electrical load to detect the load status, and requires the modification of existing electrical appliances or lines, which is difficult and costly to implement. The non-intrusive load monitoring technology analyzes the status of each load in the line by monitoring the bus. It has the advantages of strong versatility and low cost, and has become a research hotspot in recent years.

Most of the existing non-invasive load identification research is based on optimization and pattern recognition methods, but these methods all have certain defects. First, most of these models rely on a large number of labeled samples for model training, but in practical applications they often cannot obtain enough labeled data or the acquisition cost is high; second, these models usually assume that all loads in the scene are known, and new loads are added to the scene. Afterwards, the original model cannot recognize the new load, and even affects the recognition effect of the original load; finally, these models have poor generalization performance, and it is often necessary to retrain the model after changing the scene. The maintenance complexity and cost are greatly increased.

SUMMARY OF THE INVENTION

The purpose of the present invention is to propose a non-intrusive load identification method based on metric learning in view of the problems of the prior art. The load current is used as a feature, and the current feature is mapped into a new metric space through a convolutional neural network. During training, the triplet loss is used as the network loss to achieve clustering, and the distance measurement is only required for the load feature vector in the metric space during identification. This method can realize effective identification of unknown loads and has strong generalization ability; on the other hand, metric learning, as one of the methods of small sample learning, can reduce the dependence on training samples and has high practicability.

The technical scheme of the present invention is as follows:

A non-intrusive load identification method based on metric learning, comprising the following steps:

Step 1, use the event detection algorithm to monitor the power data of the power bus in real time, locate the time when the load switching event occurs, and make a difference between the steady-state data before and after the load switching event, and separate the voltage of the load switching to be identified. Current and power data;

Step 2, start a complete cycle of information collection for the current data at the voltage positive zero-crossing point, and perform resampling and normalization processing on the collected current information;

Step 3, using the trained feature extraction network to perform feature extraction on the current to obtain feature vector information of the load switching event to be identified;

Step 4, check whether the feature library exists, if not, establish a feature library, store the feature vector information and power data of the load switching event to be identified, and set the corresponding feature number;

If it exists, calculate the similarity between the eigenvector of each sample in the feature library and the eigenvector of the load to be identified one by one, and further calculate the power of the sample and the power of the load to be identified for the samples that satisfy the condition of the similarity of the eigenvectors. If the power similarity condition is met, the load switching event to be identified and the sample in the feature library belong to the same load switching event, and the feature number of the sample in the feature library is used as the feature number of the load switching event to be identified. ; If all of the eigenvector similarity conditions or power similarity conditions are not met, the load switching event to be identified is a new event, and the eigenvector information and power data of the load switching event to be identified are stored in the feature library and the corresponding features are set Numbering.

Step 5: Map the obtained feature number to an actual load switching event, where the load switching event includes load category and state transition information of the load.

Among them, the feature number of the sample set newly added to the feature library is mapped with the load switching event by the following method:

When there is a new sample added to the signature database, the user is notified, and the user can judge the actual load category of the newly added number and the state transition information of the load according to the historical switching records combined with the actual usage of the day.

The historical switching records can be used for the visual display of historical work information of various types of loads, etc.

Further, the feature extraction network consists of a plurality of one-dimensional convolutional layers, an activation function layer, a plurality of residual modules, a global average pooling layer and a linear layer, and an L2 regularization layer, which are connected in sequence.

Further, the triplet loss is used as the network loss during the training of the feature extraction network.

Further, during the training of the feature extraction network, the used training samples regard different working states of the same kind of electrical appliances as separate devices.

Further, in the step 4, the cosine similarity calculation method is used to calculate the similarity between the eigenvector of each sample in the feature library and the eigenvector of the load to be identified, and the difference calculation method is used to calculate the power of the sample and the load to be identified. Similarity of power.

Further, in the mapping between the feature number mapping and the actual load switching event, different processing is performed for the single-state load and the multi-state load, as follows:

For a single-state load, there are only two states of on/off, and its state transition is relatively simple, and only the characteristics of the state transition process need to be marked; In addition to the load type information when labeling, it is also necessary to label the change information of the load state. Only by mastering the characteristic information of the transition between the states can the multi-state load transition information and the current state information be accurately identified. .

The beneficial effects of the present invention are:

First, metric learning, as one of the small-sample learning methods, can reduce the requirement of the non-intrusive load recognition model on the number of training samples, and can obtain better recognition performance after training with a small number of samples; secondly, the load is achieved by matching with the feature library. Identify, and automatically add unknown loads to the feature library for identification and classification, and wait for later users to label device information; finally, the metric learning model has strong generalization performance, and there is no need to retrain or build a model when migrating to a new scene. It greatly reduces the migration cost and has higher practical value. At the same time, the method of the present invention also marks the current state of electrical appliances/loads in the process of establishing the mapping between the sample feature numbers newly added to the feature database and the load switching events, so that more and more accurate user-side detailed power consumption information can be obtained.

Description of drawings

Figure 1 shows the current resampling and normalization process.

FIG. 2 is a schematic diagram of the structure of the current feature extraction network.

Figure 3 is a schematic diagram of triple loss calculation.

FIG. 4 is a flow chart of the overall identification of the present invention.

FIG. 5 is a schematic diagram of multi-state load state switching.

Figure 6 shows the recognition results (confusion matrix) of the model trained in house6 in PLAID and part of the electrical test in COOLL.

Detailed ways

In order to verify the features and effects of the present invention, the present invention will be further described below in conjunction with the WHITED data set, the PLAID data set and the COOLL data set.

The invention provides a non-intrusive load identification method based on metric learning, which obtains feature vector information through a trained feature extraction network, and compares the load switching event to be identified with the feature vector information and power information of samples in a feature library. The similarity is combined with the similarity to determine whether the load switching event to be identified is a new event and to determine the event type of the load switching event to be identified that is not a new event. In this embodiment, a feature extraction network is constructed and trained first, and then the method of the present invention is described in detail in combination with the trained feature extraction network.

Build and train a feature extraction network:

(1) Training set construction:

The training set of the present invention can adopt the samples of the existing public data sets in the field of non-invasive load identification. Taking the sample of house6 in the PLAID data set as an example, the training set construction method is as follows:

House6 includes 6 kinds of electrical appliances: fluorescent lamps, air conditioners, refrigerators, fans, hair dryers, and laptop computers. Among them, air conditioners and refrigerators have 3 and 2 working states respectively. During model training, different working states are regarded as independent load types. deal with.

When collecting training samples from samples of each load type, the number of test cases for each load type is different. In order to ensure the balance of samples, when the number of samples is insufficient, the oversampling technique (synthetic minority oversampling technique, SMOTE) is used to expand them. The main process is as follows:

1) randomly select two samples x ₀ and x ₁ in the same sample;

2) The new sample is obtained according to the following formula:

x _new = x ₀ +rand(0,1)(x ₁ -x ₀ )

Rand(0,1) represents a random function that generates a random number of 0 or 1.

3) Repeat steps until the specified number of samples are obtained.

For each sample obtained, the current information is preprocessed, resampled to the specified frequency and normalized as the network input. The detailed steps are:

a In order to obtain the current phase information, the current acquisition is started when the voltage is zero-crossing in the forward direction;

b record the current information of a complete cycle;

c Perform resampling processing on the current information, and use the linear interpolation method to obtain the current information of the specified sampling frequency. The specific calculation process is as follows:

i' _n =(ceil(loc)-loc)·i _floor(loc) +(loc-floor(loc))·i _ceil(loc)

In the formula, N ₀ and N ₁ are the number of points per cycle of the original data and the data after sampling, respectively, Ts ₀ and Ts ₁ are the sampling time before and after sampling, n is the nth data point after sampling, and loc is the nth data point. Corresponding to the position of the original sequence, ceil and floor are round up and down functions respectively, i _ceil(loc) and i _floor(loc) are the nth data of the original sequence corresponding to the position of ceil(loc) and floor(loc) respectively The current value of the point, i' _n is the current value of the nth data point after sampling obtained by linear interpolation.

d Standardize the current after sampling, and scale the current amplitude to be between [-1, 1], that is, perform the following operations on the current:

In the formula, I' is the current after normalization, and I=[i' ₁ , i' ₂ ,...,i' _N1 ,] is the current sequence after sampling. max(abs(I)) represents the maximum value after taking the absolute value of the sampled current sequence.

(2) Building a feature extraction network model: the feature extraction network in the present invention can be any feature extraction network. In this implementation, reference is made to the ResNet model for construction, which is a metric network model constructed based on a one-dimensional convolution residual block. Concatenated multiple 1D convolutional layers, activation function layers, multiple residual modules, global average pooling layers, linear layers, and L2 regularization layers (not shown in the figure). Fig. 2 is a schematic diagram of an exemplary structure. The structure first uses 64 one-dimensional convolution kernels of size 7 to perform convolution operation on the current, and selects a step size of 2 for downsampling, and then passes through three residual modules in turn. Downsampling and feature extraction are performed, and finally the dimension is expanded after the global average pooling layer. After the linear layer, the feature vector extracted based on the load current is obtained. In order to eliminate the influence of the eigenvector modulus value in the distance calculation, the L2 regularization layer is used. The eigenvector output by the network is normalized so that it falls on the surface of the unit hypersphere. At this time, the Euclidean distance and the cosine distance can be regarded as equivalent. In order to speed up the convergence, a BN (BatchNormalization) layer is added to the convolutional layer and residual module after each convolution to standardize the eigenvalues. In terms of activation function selection, the network input is the current sampling value of [-1,1], so the Tanh function is selected as the activation function.

The triplet loss function is used during network training. If each pair of samples is fed into the network for training, its space complexity is O(N ³ ), and as the training process progresses, the number of triples with a loss of 0 will be More and more, the mean network loss is pulled down, making the network training slower until it falls into a local optimum. In order to solve this problem, it is necessary to filter the triples when constructing samples for each training, and select a certain number of triples whose loss is not 0 as the training sample group, as shown in Figure 3, to speed up the network training process and avoid falling into the local optimum. excellent.

Specifically, the triplet loss is used as the loss function for network training, and the triplet loss is as follows:

L=max(d(a,p)-d(a,n)+margin,0)

In the formula, the triplet is <a,p,n>, a is called the anchor, p is a positive instance (positive), which belongs to the same category as a, n is a negative example (negative), and a different classes. d is the distance metric function, and the Euclidean distance is used during training (due to normalization, the farther the distance on the unit hypersphere is, the larger the angle between the two vectors is, which is equivalent to the cosine distance), and the margin is A constant whose purpose is to prevent the loss from being 0 when both positive and negative samples are very close to the anchor samples.

After the training is completed, a trained feature extraction network is obtained, and the trained feature extraction network is utilized to realize the non-intrusive load identification method based on metric learning of the present invention, as shown in Figure 4, comprising the following steps:

Step 1, use the event detection algorithm to monitor the power bus power data in real time, locate the occurrence time of the load switching event, and make a difference between the steady-state data before and after the occurrence time of the load switching event, and isolate the load switching event to be identified. Voltage, current and power data;

Step 2, starting a complete cycle of current information collection for the voltage and current data at the zero-crossing point of the forward voltage, and re-sampling and normalizing the collected current information;

Step 3, using the trained feature extraction network to perform feature extraction on the current to obtain feature vector information of the load switching event to be identified;

Step 4, check whether the feature library exists, if not, establish a feature library, store the feature vector information and power data of the load switching event to be identified, and set the corresponding feature number;

If it exists, calculate the similarity between the eigenvector of each sample in the feature library and the eigenvector of the load to be identified one by one, and further calculate the power of the sample and the power of the load to be identified for the samples that satisfy the condition of the similarity of the eigenvectors. If the power similarity condition is met, the load switching event to be identified and the sample in the feature library belong to the same load switching event, and the feature number of the sample in the feature library is used as the feature number of the load switching event to be identified. ; If all of them do not meet the feature vector similarity condition or power similarity condition, the load switching event to be identified is a new event, and the feature vector information and power data of the load switching event to be identified are stored in the feature library and the corresponding feature number is set .

The similarity calculation result is used to judge the similarity, and calculation methods such as cosine similarity, difference calculation, and Euclidean distance can be used. Preferably, the present embodiment adopts the cosine similarity calculation method to calculate the feature vector of each sample in the feature library. The similarity with the feature vector of the load to be identified is as follows:

In the formula, x and y represent two eigenvectors, and n represents the length of the eigenvectors. The closer the value is to 1, the more similar the two features are. In practice, the similarity condition can be set reasonably according to the situation.

The difference calculation method is used to calculate the similarity between the power of the sample and the power of the load to be identified, as follows:

Relative difference = (P _max -P _min )/P _min

Among them, P _max and P _min are the large value and the small value of the power of the load switching event to be identified and the power of the samples in the feature library, respectively.

In practice, similarity conditions can be reasonably set according to the situation, and different thresholds can also be set for different power segments to improve the recognition accuracy.

Among them, the feature number of the sample set newly added to the feature library is mapped with the load switching event by the following method:

When there is a new sample added to the feature database, the user is notified, and the user can judge the actual load category of the newly added feature number and the state transition information of the load according to the historical switching records combined with the actual usage of the day. Among them, for a single-state load, there are only two states of on/off, and the state transition is relatively simple, and only the characteristics of the state transition process need to be marked; for a multi-state load, as shown in Figure 5, due to the various states of the load There may be transitions between them, and different states have different characteristic information. In addition to the load type information, the change information of the load state can also be annotated, and the characteristic information of the transition between the states can be grasped. Accurately identify the multi-state load conversion information and the current state information, so as to obtain more and more accurate detailed power consumption information on the user side.

To illustrate the advantages of the method of the present invention, two examples are constructed for illustration.

In the two embodiments, in terms of model parameter selection, the extracted current sampling frequency is 128 points per cycle, the feature dimension extracted by the metric network is 16, and the similarity threshold is set to 0.8 when the cosine similarity is judged. When the cosine similarity is greater than 0.8, the two features are considered to be similar, otherwise they are not similar. Similarly, when performing power matching, set the power relative difference threshold to 0.2.

In terms of test result measurement indicators, the commonly used indicators of F ₁ score and accuracy (Acc) are selected for measurement, which are calculated as follows:

In the formula, TP, FP, TN, and FN refer to the number of true cases, false positive cases, true negative cases, and false negative cases, respectively, precision is the accuracy rate, and recall is the recall rate.

Example 1:

In order to illustrate the general recognition ability of the present invention, 80% of the WHITED data set samples are used as the training set to train the feature extraction network, and the rest of the samples are tested, which is similar to the common VI method: Reference [1] (De Baets L, Dhaene T, Deschrijver D , et al.VI-Based Appliance Classification Using Aggregated PowerConsumption Data[C]//2018 IEEE International Conference on Smart Computing (SMARTCOMP). Taormina: IEEE, 2018: 179–186.), literature [2] (Wang Ying, Yang Wei, Xiao Xianyong, Zhang Shu. Non-intrusive residential load monitoring method based on UI trajectory curve refined identification [J]. Power Grid Technology, 2021, 45(10): 4104-4113.) and literature [3] (De Baets L , Ruyssinck J, Develder C, et al.ApplianceClassification Using VI Trajectories and Convolutional Neural Networks[J].Energy and Buildings,2018,158:32–36.) for comparison, using the F ₁ score as the performance indicator, the results are shown in the table 1 shown.

Table 1 Comparison of general recognition capabilities

It can be seen that the present invention can obtain better identification ability than the existing V-I method (including V-I trajectory and power two-stage identification) by using current data and power data in combination with a one-dimensional convolutional neural network. The method of the present invention has Simpler network structure and less computational complexity.

Example 2:

In order to verify the generality and small sample learning ability of the method of the present invention, only the samples of house6 in the PLAID data set are used as the training set, and some electrical appliances in the COOLL data set are selected for testing.

The electrical appliances selected in the COOLL data set are air conditioners, modems, chargers, travel chargers, drilling machines, fan 1, fan 2, soldering iron, and 9 types of vacuum cleaners. 20 sets of information are collected for each electrical appliance as test samples, and the identification results are obtained. The confusion matrix is shown in Figure 6.

Due to signal noise or current fluctuations during steady-state operation of electrical appliances, some electrical appliances are identified as multiple numbers. As shown in the confusion matrix, electrical appliances 1 and 5 are respectively identified with two numbers, that is, the intra-class waveform difference exceeds Thresholds set by the model, some samples are identified as new devices. However, the overall recognition effect is ideal. The recognition accuracy of each electrical appliance is shown in Table 2. Electrical appliances that are identified as multiple numbers can be marked by mapping the feature numbers set in the samples added to the feature library to the actual load switching event. , mapping these numbers to the same type of electrical appliances to solve, the method of the present invention has good robustness.

Table 2 Cross-dataset generalization ability test results

To sum up, it can be seen from the examples that, compared with other methods based on V-I trajectory, the present invention is better in general recognition performance, and more importantly, the present invention has the ability to generalize across datasets and small sample learning. The performance is still excellent in the test. It only needs to be trained with limited samples (only the load in house6 in the PLAID data set is used as the training set), and in the later stage, multiple identification numbers can be mapped to the same electrical appliance to further improve the identification effect. At the same time, the method of the present invention also marks the status of the current electrical appliance/load in the process of establishing the mapping between the sample feature number newly added to the feature database and the load switching event, so that more and more accurate detailed power consumption on the user side can be obtained. information.

Obviously, the above-mentioned embodiments are only examples for clear description, and are not intended to limit the implementation manner. For those of ordinary skill in the art, changes or modifications in other different forms can also be made on the basis of the above description. All implementations need not and cannot be exhaustive here. However, the obvious changes or changes derived from this are still within the protection scope of the present invention.

Claims (5)

1. A non-intrusive load identification method based on metric learning is characterized by comprising the following steps:

step 1, monitoring power bus power data in real time by using an event detection algorithm, positioning the occurrence time of a load switching event, and differentiating steady-state data before and after the occurrence time of the load switching event to separate out voltage and current and power data of switching of a load to be identified;

step 2, starting information acquisition of a complete period of the current data at the voltage positive zero crossing point, and performing resampling and normalization processing on the acquired current information;

step 3, using the trained feature extraction network to extract the features of the current to obtain the feature vector information of the load switching event to be identified;

step 4, checking whether the feature library exists or not, if not, establishing the feature library, storing feature vector information and power data of the load switching event to be identified and setting a corresponding feature number;

if the load switching event is the same as the sample in the feature library, the feature number of the sample in the feature library is used as the feature number of the load switching event to be identified; and if all the characteristic vector similarity conditions or the power similarity conditions are not met, the load switching event to be identified is a new event, the characteristic vector information and the power data of the load switching event to be identified are stored in a characteristic library, and corresponding characteristic numbers are set.

And 5, mapping the obtained feature numbers into actual load switching events, wherein the load switching events comprise load types and load state conversion information.

The feature number of the sample setting newly added into the feature library is mapped with the load switching event by the following method:

and when a sample newly added into the feature library exists, informing a user, and judging the actual load type of the newly added number and the state conversion information of the load by the user according to the historical switching record and the current-day actual use condition.

2. The non-invasive load identification method based on metric learning as claimed in claim 1, wherein the feature extraction network is composed of a plurality of one-dimensional convolution layers, an activation function layer, a plurality of residual modules, a global average pooling layer and a linear layer, and an L2 regularization layer, which are connected in sequence.

3. The non-intrusive load identification method based on metric learning as claimed in claim 1, wherein the feature extraction network training uses triplet loss as network loss.

4. The non-intrusive load identification method based on metric learning as claimed in claim 1, wherein in the feature extraction network training, training samples used in the training process regard different working states of the same electrical appliance as separate devices.

5. The non-invasive load identification method based on metric learning as claimed in claim 1, wherein in the step 4, the cosine similarity calculation method is used to calculate the similarity between the feature vector of each sample in the feature library and the feature vector of the load to be identified, and the difference calculation method is used to calculate the similarity between the power of the sample and the power of the load to be identified.

Images