CN114818827A

CN114818827A - Non-invasive load decomposition method based on seq2point network

Info

Publication number: CN114818827A
Application number: CN202210546135.2A
Authority: CN
Inventors: 陆子刚; 黄奇峰; 卢树峰; 纪峰; 孙永辉; 张亦苏; 左强; 王忠东; 徐敏锐; 陈刚; 欧阳曾恺; 吴桥
Original assignee: State Grid Jiangsu Electric Power Co ltd Marketing Service Center; State Grid Corp of China SGCC; State Grid Jiangsu Electric Power Co Ltd; Hohai University HHU
Current assignee: State Grid Jiangsu Electric Power Co ltd Marketing Service Center; State Grid Corp of China SGCC; State Grid Jiangsu Electric Power Co Ltd; Hohai University HHU
Priority date: 2022-05-19
Filing date: 2022-05-19
Publication date: 2022-07-29

Abstract

The invention discloses a non-invasive load decomposition method based on a seq2point network, which comprises the steps of constructing a seq2point non-invasive load decomposition model and training the model; reading the time sequence of the total load power by a sliding window to generate an input sequence; inputting the input sequence into the one-dimensional convolution layer to improve the one-dimensional convolution network to automatically extract the characteristics of the input sequence and obtain the distributed characteristics of the power data; and placing the extracted distributed power characteristics in a full connection layer with a fixed length for storage, outputting the distributed characteristics of the power data integrated into the sample space through an activation function, and obtaining a decomposed power sequence, thereby realizing the Seq2point load decomposition. The method fully considers the data feature extraction and the data feature based on the time sequence, and performs feature self-extraction on the data through Conv1D, so that the identification error under low-frequency sampling is reduced. The invention has better generalization capability and can identify various electric appliances.

Description

Non-invasive load decomposition method based on seq2point network

Technical Field

The invention relates to a non-intrusive load identification technology, in particular to a non-intrusive load decomposition method based on a seq2point network.

Background

After the policy of carbon peak reaching and carbon neutralization is provided, the national power grid takes carbon neutralization as the ultimate goal, the construction of a clean energy internet optimization configuration platform is accelerated, and the fusion innovation and the application of technologies such as big data, cloud computing, the internet of things and artificial intelligence in the aspects of electric power, energy and the like are enhanced. In this context, research on non-invasive monitoring techniques for consumer appliances has become far-reaching. On one hand, the non-invasive research on the electric appliances of the user can reduce the energy consumption of the user, for example, the using condition of the electric appliances is monitored, unnecessary no-load electric appliances are closed, and therefore the purposes of reducing the electric energy waste and reducing the carbon emission are achieved. On the other hand, the data of the non-invasive load monitoring research can provide solid data support for large data research electric power use conditions and the like.

Since the concept of Non-invasive Load Monitoring was proposed by professor Hart of massachusetts, Non-invasive Load Monitoring (NILM) has received much attention. In 2015, Kelly applies deep learning to the NILM problem for the first time, and builds three different deep learning network frames based on a recurrent neural network, an automatic denoising encoder and a deep neural network respectively, so that the deep learning has good performance in processing the NILM problem. The document [ a load decomposition method based on a one-dimensional convolutional neural network ] proposes a Sequence-to-Point (Seq 2Point) network, and proves that under the same network system, the result output by the Seq2Point network is closer to target distribution than the Sequence-to-Sequence (Seq 2Seq) network; the input sequence characteristics are obtained based on a sliding window and a sequence expansion module, and the load decomposition method based on a one-dimensional Convolutional Neural Network (CNN) is proved to have superiority on a UK _ DALE data set through experiments.

The literature [ CNN-based non-invasive monitoring technology of the household electrical appliances ] adopts a non-noise current component as an input characteristic for training a convolutional neural network, and the final example proves the effectiveness of the method. The literature [ non-invasive load identification based on a 1D-CNN model ] designs a one-dimensional convolution neural network which can be used for low-frequency data and high-frequency data. The document [ non-invasive load decomposition algorithm based on seq2seq Model ] proposes a load recognition algorithm from sequence to sequence based on CNN and LSTM (Long Short-term Memory) to perform Model training and testing by using REDD data set, and compares the algorithm with algorithms such as FHMM (factory Hidden Markov Model), CO (Combinatorial Optimization) and the like, so as to obviously improve the result. In the literature [ study on household appliance load identification method based on clustering features and seq2seq depth CNN ] firstly, the running states of electric equipment are extracted by using improved iterative K-means clustering to establish load feature sets, and then the feature sets are input into a constructed sequence-to-sequence one-dimensional deep convolutional neural network model and a sequence-to-sequence single-direction and two-direction long-time memory network model to carry out load decomposition and excavation on the running states of the equipment.

The non-invasive load identification at the present stage has a series of problems, such as poor data quality and low identification accuracy. Firstly, data collection is abnormal due to sensor faults, and poor data quality is caused by the fact that the number of samples of the electric appliance in the running state is small. For machine learning models, abnormal data will cause the model to learn the wrong input-output correspondence, reducing the recognition accuracy of the model. And less operation samples make the model tend to learn the state of the electric appliance as non-operation, so that it is difficult to train an excellent learning model. Secondly, the traditional machine learning method depends on a priori parameter selection method, and often results in inappropriate parameter selection, incapability of fitting real data by an algorithm and the like, so that the non-invasive load identification accuracy is low, and the high accuracy requirement of an actual scene is difficult to meet.

Disclosure of Invention

In order to solve the defects in the prior art, the invention aims to provide a non-intrusive load decomposition method based on a Seq2point network, which reads a time sequence of a total load by a sliding window to generate an input sequence so as to improve the automatic extraction of input sequence characteristics by a one-dimensional convolution network, thereby realizing the decomposition of the Seq2point load, improving the performance of the load decomposition and obtaining higher decomposition precision.

The invention adopts the following technical scheme.

A non-intrusive load decomposition method based on a seq2point network comprises the following steps:

(1) constructing a seq2point non-invasive load decomposition model and training the model;

(2) reading the time sequence of the total load power by a sliding window to generate an input sequence;

(3) inputting the input sequence into the one-dimensional convolution layer to improve the one-dimensional convolution network to automatically extract the characteristics of the input sequence and obtain the distributed characteristics of the power data;

(4) and placing the extracted distributed power characteristics in a full connection layer with a fixed length for storage, outputting the distributed characteristics of the power data integrated into the sample space through an activation function, and obtaining a decomposed power sequence, thereby realizing the Seq2point load decomposition.

Further, in the step (1), the seq2point non-intrusive load decomposition model comprises a BN layer, a one-dimensional convolution layer, a one-dimensional pooling layer, a full connection layer and an activation function.

Further, in the step (1), a mean square error MSE is adopted as a loss function to train the model, and error calculation is carried out on each training result and the weight of the model is adjusted.

Further, in the step (1), model tuning is performed by using an Adam optimizer.

Further, in the step (2), the input sequence is normalized by a BN layer, the mean value of the same batch of data is normalized to 0, and the variance is normalized to 1.

Further, the step (4) further comprises the step of performing maximal pooling on the extracted distributed power characteristics through a one-dimensional pooling layer, and then storing the distributed power characteristics in a full-connection layer with a fixed length.

Further, in step (4), an activation function ReLU is employed.

Further, in step (1), the loss function MSE is the square of the difference between the real value and the predicted value, and then the sum is averaged, and the calculation method is as follows:

for non-intrusive load identification tasks, it is trueThe real label should be a power vector of 1 × n, denoted as y; the model output is recorded as

Further, in the step (2), the calculation mode of normalization processing of the BN layer:

Y＝(X-μ(X))/σ(X)

wherein, X is the input data sequence of the same batch, Y is the data sequence after normalization, and mu (X) and sigma (X) are the mean value and standard deviation of the X sequence respectively.

Further, in the step (4), the full connection layer is a classifier in the convolutional neural network, and the feature vectors after the pooling operation are classified by the classifier to obtain the required output.

Compared with the prior art, the non-intrusive load decomposition method based on the Seq2point network has the advantages that the time sequence of the total load is read by the sliding window to generate the input sequence, so that the one-dimensional convolution network is improved to automatically extract the characteristics of the input sequence, the Seq2point load decomposition is realized, the load decomposition performance is improved, and the higher decomposition precision is obtained. The experimental results on the disclosed data set REDD show that the load decomposition method based on the one-dimensional convolutional neural network provided by the invention has better effect and better performance compared with other algorithms.

The algorithm provided by the invention has a simple network structure and strong functions, fully considers the characteristics of data feature extraction and data based on time sequence, adopts an Adam optimizer for optimization, and obtains better effect by respectively comparing the REDD data set with other three non-invasive load decomposition algorithms.

According to the method, the low-frequency data in the REDD data set is used for carrying out experiments, the characteristics of the data are self-extracted through Conv1D, the traditional electric appliance characteristics which can be obtained only by using the high-frequency data are avoided, and the identification error under the low-frequency sampling is reduced.

The seq2point load decomposition algorithm has good generalization capability, and can be used for identifying various electrical appliances by using the multi-day data of the various electrical appliances for verification.

Drawings

FIG. 1 is a flow chart of a non-intrusive load decomposition method based on a seq2point network according to the present invention;

FIG. 2 is a schematic diagram of a seq2point non-intrusive load decomposition model;

FIG. 3 is an example of the operation of a one-dimensional convolutional layer (Conv 1D);

fig. 4 is an example of a max pooling operation with a pool size of 3.

Detailed Description

The present application is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present application is not limited thereby.

The non-invasive load identification has important significance on power scheduling, risk estimation and the like of a power grid system. Aiming at the problems that the algorithm recognition accuracy of the non-invasive load decomposition at the present stage is low and the operation cost is high, the invention provides a non-invasive load decomposition algorithm based on a seq2point network, and the non-invasive load decomposition algorithm is based on a sequence-to-point deep learning model of a convolutional neural network, and the non-invasive load decomposition algorithm can have better load recognition capability on household appliances under the condition of low-frequency sampling of 1Hz or below. The disclosed low-frequency data set is used for training and testing on the model, and compared with the existing algorithm, the model obtains higher accuracy on the low-frequency data set, meanwhile, the algorithm complexity is reduced, and the method has obvious advantages.

As shown in fig. 1, the non-intrusive load decomposition method based on the seq2point network includes the steps of:

(1) constructing a seq2point non-invasive load decomposition model, which comprises a BN layer, a one-dimensional convolution layer, a one-dimensional pooling layer, a full-link layer and an activation function, and is shown in FIG. 2; training the model;

in the model training process, a loss function needs to be set, error calculation is carried out on each training result, and the model weight is adjusted. For the non-invasive load identification task, the real label of the non-invasive load identification task is a power vector of 1 multiplied by n and is marked as y; the model output is recorded as

The invention adopts MSE (Mean Square Error) as a loss function, wherein the MSE is the Square of the difference value between the real value and the predicted value and then the sum is averaged. The Loss function Loss calculation method comprises the following steps:

the key of model training is model parameter adjustment, and optimization of model parameters has important influence on improvement of model performance. The Adam optimizer is used, the Adam optimizer can replace a traditional random gradient descent (SGD) optimization algorithm, the traditional SGD is low in convergence speed and difficult to select appropriate learning efficiency, the AdaGrad optimizer combines the advantages of AdaGrad and RMSProp, each parameter uses the same learning efficiency, the learning efficiency can be automatically updated, efficient calculation can be performed by using less memory, and the Adam optimizer is suitable for the optimization problem of large-scale data. Moreover, Adam is a momentum-based algorithm that uses historical information of the gradient, which need hardly be parametrized.

specifically, the input sequence is normalized by the BN layer.

Bn (batch normalization) is to normalize the data of the same batch, normalize the mean value of the data of the same batch to 0, normalize the variance to 1, and specifically calculate the following formula:

Y＝(X-μ(X))/σ(X) (2)

The input X with different distributions can be normalized to Y with the mean value of 0 and the variance (standard deviation) of 1 through the calculation of the formula (2), so that the gradient of the flowing network can be improved, the learning rate is allowed to be higher, the training speed is accelerated, and the generalization capability of the network is improved.

(3) Then inputting the input sequence into the one-dimensional convolution layer to improve the one-dimensional convolution network to automatically extract the characteristics of the input sequence and obtain the distributed characteristics of the power data;

the data sampled at low frequency does not have the characteristics of transient state and steady state, the CNN is introduced for mining the information in the data to perform characteristic self-extraction on the data, and most of the data at low frequency only collects the attribute value of power, such as REDD data set and UK-DALE data set, so that the time relationship hidden in the data is particularly important.

Because the actual operation of the electric appliance has an unstable state, the power measurement also has errors, a large amount of nonlinear disturbance is caused, and the actual operation state of the electric appliance is difficult to fit by directly using a linear model. Therefore, a non-linear one-dimensional convolutional layer (Conv1D) is used to convolve the power sequence on a one-dimensional scale, and the power features are extracted by means of convolution kernels, and an example of a typical Conv1D is shown in fig. 3. The use of Conv1D avoids the traditional manual feature extraction and the structure is robust.

And taking the total load power time sequence as an input vector, and performing convolution operation on the input vector by using a convolution core to obtain the distributed characteristics of the power data.

(4) And placing the extracted distributed power characteristics in a full connection layer with a fixed length for storage, outputting the distributed characteristics of the power data integrated into the sample space through an activation function ReLU, and obtaining a decomposed power sequence, thereby realizing the Seq2point load decomposition.

Specifically, the extracted distributed power features are subjected to maximum pooling through a one-dimensional pooling layer, and then are stored in a full-connection layer with a fixed length.

For one-dimensional data, the pooling layer is a one-dimensional pooling, including a one-dimensional maximum pooling, a one-dimensional mean pooling, and the like. The invention adopts the maximum pooling. FIG. 4 is an example of a typical pooling operation with a pool size of 3. The pooling layer can reduce data characteristics on the premise of keeping the scale invariance of data, reduce parameters and calculated amount, prevent overfitting and improve the generalization capability of the model.

The fully-connected layer is a classifier in a convolutional neural network. After convolution and pooling operations are carried out on the original data, the original data are mapped to a feature space of a hidden layer, and feature vectors are extracted. And then classifying the feature vectors by a classifier to obtain the required output.

The fully-connected layer includes a weight W and an offset B, taking one-dimensional fully-connected layer calculation as an example, let a one-dimensional input be X, a weight matrix be W, and an offset be B, then an output Y can be calculated by the following formula:

Y＝W ^T X+B (3)

by adjusting the sizes of W and B, if W is adjusted to 2 × 3 and B is changed to 2 × 1, an output Y with a size of 2 × 1 can be obtained. For this research task, the output size should be 1 × n, where n is the number of appliances.

The data set used experimentally in the present invention is the REDD data set, the first publicly available data set specifically collected to aid NILM studies. Which contains data of weeks of power supply for 6 different premises and data of high frequency current/voltage of the main power supply of two of the premises. The high-frequency data is often used for traditional transient characteristic extraction, so the low-frequency data is used for experiments, the sampling frequency of the main power data is 1s once, and the sampling frequency of the sub-table is 3s/4s once.

In order to verify the decomposition capability and generalization performance of the model, 4 electric appliances with the front power are selected to train on a training set aiming at different decomposition methods, and then evaluation indexes of different decomposition methods are calculated according to test results of a test set.

The Root Mean Square Error (RMSE) evaluation model is adopted, and the calculation method is as follows:

wherein i is the number of the electric appliance, m is the total number of the test samples, and y _i,k For the true value of the instrument i in the k samples,

is the output value of the electric appliance i in the decomposition model pair k samples.

The experimental results are shown below, and Table 1 shows the comparison between the method of the present invention and the experimental results of CO, FHMM and the method of the literature [ Sequence-to-point learning with neural networks for non-invasive load monitoring ]. Table 2 shows the number of parameters of the process of the invention compared to the literature process.

TABLE 1

TABLE 2

Compared with the traditional method CO and FHMM, the algorithm of the invention performs regression prediction on data based on deep learning, and performs deep mining on the data through a multilayer network, so that the performance of the algorithm is superior, and the RSME is reduced by 65%.

Compared with a literature method, the algorithm greatly reduces the number of parameters, has higher operation speed, and simultaneously improves the decomposition effects of different electrical appliances to different degrees by virtue of the Seq2 point. The algorithm of the invention can have better load identification capability for household appliances under the low-frequency sampling of 1Hz and below.

The present applicant has described and illustrated embodiments of the present invention in detail with reference to the accompanying drawings, but it should be understood by those skilled in the art that the above embodiments are merely preferred embodiments of the present invention, and the detailed description is only for the purpose of helping the reader to better understand the spirit of the present invention, and not for limiting the scope of the present invention, and on the contrary, any improvement or modification made based on the spirit of the present invention should fall within the scope of the present invention.

Claims

1. A non-intrusive load decomposition method based on a seq2point network is characterized in that,

the method comprises the following steps:

2. The method of claim 1, wherein the method comprises the steps of,

in the step (1), the seq2point non-invasive load decomposition model comprises a BN layer, a one-dimensional convolution layer, a one-dimensional pooling layer, a full-link layer and an activation function.

3. The method of claim 1, wherein the method comprises the steps of,

in the step (1), a mean square error MSE is used as a loss function to train the model, and error calculation is carried out on each training result and the weight of the model is adjusted.

4. The method of claim 1, wherein the method comprises the steps of,

in the step (1), model parameter adjustment is carried out by using an Adam optimizer.

5. The method of claim 1, wherein the method comprises the steps of,

in the step (2), normalization processing is further performed on the input sequence through a BN layer, the mean value of the data in the same batch is normalized to 0, and the variance is normalized to 1.

6. The method of claim 1, wherein the method comprises the steps of,

and (4) performing maximum pooling on the extracted distributed power characteristics through a one-dimensional pooling layer, and storing the distributed power characteristics in a full-connection layer with a fixed length.

7. The method of claim 1, wherein the method comprises the steps of,

in step (4), an activation function ReLU is used.

8. The method of claim 3, wherein the method of non-intrusive load decomposition based on seq2point network is characterized in that,

in the step (1), a loss function MSE is the square of the difference value between a real value and a predicted value, and then the sum is averaged, and the calculation method comprises the following steps:

for the non-invasive load identification task, the real label of the non-invasive load identification task is a power vector of 1 multiplied by n and is marked as y; the model output is recorded as

9. The method of claim 5, wherein the method comprises the steps of,

in the step (2), the BN layer normalization processing calculation mode is as follows:

Y＝(X-μ(X))/σ(X)

10. The method of claim 6, wherein the method of non-intrusive load decomposition based on seq2point network is characterized in that,

in the step (4), the full connection layer is a classifier in the convolutional neural network, and the feature vectors after the pooling operation are classified by the classifier to obtain the required output.