CN115392318A - Intelligent building non-invasive load decomposition method based on parallel connection network - Google Patents

Abstract

The invention provides a non-invasive load decomposition method based on a parallel connection network. Then, a parallel connection network model is constructed, features are extracted through a parallel connection network combining a cavity residual convolution neural network and a bidirectional long-short term memory network, the characterization capability of the features is improved, an attention mechanism is introduced, redundant information is eliminated, important information is further extracted, and the decomposition effect of the model is improved. And then, training the decomposition model by using the training data to obtain an optimized decomposition model. And finally, inputting the actually measured original data into the optimized decomposition model for load decomposition to obtain a decomposition result, and analyzing the experimental result.

Description

Intelligent building non-invasive load decomposition method based on parallel connection network

Technical Field

The invention relates to an intelligent building non-invasive load decomposition method based on a parallel connection network, which abandons the traditional layer-by-layer connection decomposition model, adopts a hollow residual convolution neural network and a bidirectional long-short term memory network for parallel connection, deeply extracts load power characteristics, improves the representation capability of non-invasive load decomposition, further extracts important characteristics by an attention mechanism, screens out useless characteristics, reduces model parameters, improves the decomposition efficiency and the decomposition precision of the model, has strong universality and lower cost of extracted power characteristics, can be widely applied to intelligent building construction, constructs a nationwide intelligent building non-invasive load monitoring system, and realizes nationwide intelligent power utilization.

Background

With the rapid development of smart grids, various power enterprises are transformed to intensification and technology, and a more reliable, stable and efficient power system is needed. This requires that the specific information in the power generation, transmission and utilization processes of the power system be fully grasped and controlled. The rapid improvement of the intelligent power utilization technology of the whole power grid and the rapid improvement of the power utilization efficiency and the economy become decisive factors for constructing the intelligent power grid. Therefore, how to realize flexible and interactive intelligent power utilization and improve the power utilization efficiency of the user side becomes a hot topic of current research.

For constructing a nationwide intelligent power utilization system, the method realizes the benign interaction between a user side and a power grid side, and the key step is load monitoring. Only if the detailed power utilization information of the user is mastered, scientific power utilization of the user can be better guided, and reference is made for formulating a policy for adjusting the power utilization mode of the demand side, so that the purposes of saving energy and optimizing power dispatching are achieved. The main load monitoring methods at present comprise an invasive load monitoring method and a non-invasive load monitoring method, the invasive load monitoring method needs to monitor each power load in real time by installing a monitoring device, the monitoring mode has high data acquisition reduction degree and precision, but the installation and maintenance cost of the device is high; in addition, the privacy of residents may be revealed in each load installation monitoring, and most residents may have a conflict psychology. The Non-invasive Load monitoring is also called Non-invasive Load differentiation (NILD), under the premise of not changing the existing circuit structure of a household, a smart electric meter is additionally arranged at a household end to obtain household electricity data (voltage, current, power and the like), and the decomposition result is obtained through a Load decomposition algorithm, so that each Load state in the current system is obtained. In order to protect the power consumption privacy of the user, under the condition of meeting the power consumption comfort of the signed user, the non-invasive load decomposition is carried out on the high-power controllable load of the signed user, the user is helped to actively participate in demand response to the maximum extent, and the benefit maximization is realized.

The hardware part of the non-invasive load monitoring is an intelligent measuring device located at the client side. As shown in fig. 1, the device monitors the power utilization conditions of various electric devices under the device. The electric quantities of the total load, such as voltage, current and the like, measured by the device can be regarded as signals bearing electric power information and comprise load component information with different characteristics, and the NILD system can realize load decomposition by extracting the characteristic information of the electric quantities and estimate the energy utilization information of the single electric equipment, such as the use state, the energy consumption and the like.

The traditional decomposition model adopts a layer-by-layer connection mode to extract features, gradient disappearance and overfitting can occur along with depth deepening, important information can be lost in layer-by-layer feature extraction, and the decomposition performance of the model is reduced. In addition, with the diversification of the types of the electric loads, the load characteristic characterization capability extracted by a single characteristic extraction model is insufficient, so that the accuracy of load decomposition is reduced.

Disclosure of Invention

According to the existing problems, in order to solve the problems of overfitting and gradient disappearance caused by a deep network and the problem of information disappearance caused by extracting features layer by layer, the invention provides an intelligent building non-invasive load decomposition method based on a parallel connection network. In addition, the invention uses the cavity convolution and the residual connection to increase the receptive field and solve the problem of gradient disappearance.

Step1: selecting a part of high-power controllable loads of intelligent building subscribers as research objects, collecting total power data and power data of each target electrical appliance by using a non-invasive load decomposition measuring device, and preprocessing the data. Further processing is performed using the measured data, generating a model training data set, and calculating a mean value of the measured data set at 8:2, the data set is divided into a training set and a test set.

And 2, step: and constructing a decomposition model based on the parallel connection network.

And 3, step3: and training and testing the model by using the training set and the testing set. The decomposition model achieves the best decomposition effect by continuously adjusting the model parameters and the network structure. And inputting the actually measured original data into the optimized decomposition model for load decomposition to obtain the decomposition result of each target electrical appliance.

And 4, step4: and analyzing the decomposition effect graph and the parameter index.

The invention provides a non-invasive load decomposition method based on a parallel connection network. Then, a parallel connection network model is constructed, features are extracted through a parallel connection network combining a cavity residual convolution neural network and a bidirectional long-short term memory network, the characterization capability of the features is improved, an attention mechanism is introduced, redundant information is eliminated, important information is further extracted, and the decomposition effect of the model is improved. And then, training the decomposition model by using the training data to obtain an optimized decomposition model. And finally, inputting the actually measured original data into the optimized decomposition model for load decomposition to obtain a decomposition result, and analyzing the experimental result.

Drawings

In order to provide a clearer understanding of embodiments of the present patent, reference will now be made briefly to the following drawings, which illustrate embodiments of the present patent:

FIG. 1 is a diagram of a non-invasive load monitoring and measuring apparatus;

FIG. 2 is a diagram of a non-intrusive load resolution network architecture;

FIG. 3 is a detailed process diagram of load resolution;

Detailed Description

The present invention will be explained in further detail by a specific example.

In this embodiment, the intelligent building non-intrusive load decomposition method based on the parallel connection network specifically includes the following steps:

step1: selecting a part of high-power controllable loads of intelligent building subscribers as research objects, collecting total power data and power data of each target electrical appliance by using a non-invasive load decomposition measuring device, and preprocessing the data. Further, data construction is performed using the measured data, a model training data set is generated, and the ratio of 8: the ratio of 2 divides the data set into a training set and a testing set. The method comprises the following specific steps: the data measurement is the first step of the non-invasive load technology, the total current, the total voltage and the total power of the resident household are firstly obtained through a total electric meter, then the characteristics of harmonic waves, waveform distortion and the like of the resident household are calculated in real time in a segmented mode, and frequency domain transformation and wavelet change analysis can also be carried out. The intelligent electric meters with different sampling frequencies need to be selected according to different load characteristics. Because a large amount of interference and errors exist in the data measurement process, preprocessing needs to be performed by means of mean filtering, wavelet denoising and the like. The method is characterized in that partial high-power electric appliances of contracted residents are selected as research objects, such as air conditioners, water heaters, refrigerators, induction cookers, microwave ovens, washing machines, disinfection cabinets, floor sweeping robots, electric cookers and the like. And (3) acquiring an active power sequence and a total power sequence of a single device by using a measuring device, preprocessing data, and then normalizing by adopting a maximum and minimum normalization method to obtain usable measured data. And further processing using these measured data, generating a large amount of training data, and calculating a mean value of the measured data according to 8:2 into a training set and a test set.

Step2: and constructing a decomposition model based on the parallel connection network. As shown in fig. 2, a structure diagram of a non-invasive load decomposition network is obtained by firstly performing data processing on input total power and then inputting the processed input total power into a cavity residual error convolution network and a bidirectional long-short term memory network for feature extraction. In order to realize multi-angle feature extraction, the traditional layer-by-layer connection mode is abandoned, and the spatial features and the time sequence features of input data are respectively extracted by adopting parallel connection; then, synchronous processing is carried out on different characteristics to obtain characteristic data with the same dimensionality. And then, the feature data with the same dimension are connected in series to obtain feature data with composite features. And finally, inputting the composite features into an attention mechanism module, further extracting the features, improving the characterization capability of the features, and performing decomposition output through a full connection layer. The decomposition model can be divided into a cavity residual convolution module, a bidirectional long-short term memory network module and an attention mechanism module, and has the following specific structure:

1) Hollow residual convolution module

A Convolutional Neural Network (CNN) extracts features of input data through a plurality of filters. The filters perform layer-by-layer convolution and pooling on input data, and extract topological structure features contained in the input data layer by layer. Although the feature extraction effect of the convolutional neural network is very strong, as the number of network layers increases, problems of gradient disappearance, overfitting and the like occur. In order to make up for the defects of insufficient load characteristic utilization rate, data loss, gradient disappearance and the like, residual error connection is introduced, the idea of cross-layer connection is used for reference, the original input feature mapping and the output feature mapping of the rear layer are added, the feature fusion is completed by using an activation function, and the residual error connection enables the front and rear information transfer of the network to be smoother. Defining the bottom layer mapping as H (x), adding the input x of the residual unit, and fitting a residual function F (x) = H (x) + x through the stack layer to obtain the output. The residual join consists mainly of two weight layers, each containing weight parameters. x represents the input and the residual join expression is as follows:

F(x)＝ω ₂ σ(ω ₁ x) (1)

where σ denotes the nonlinear activation function of ReLU, ω ₁ ,ω ₂ Is a weight parameter. Most preferablyAnd then adding the input end x and the output result of the second weight layer in a cross-layer connection transfer mode to obtain the final output of the whole residual error structure. The calculation formula is as follows:

y＝F(x,{ω _i })+x (2)

wherein, F (x, { ω _i }) is the residual mapping function to be trained and learned in residual concatenation, { omega } _i Is the set of weight parameters. Input feature map x connected by the 1 st residual in the residual structure according to equations (1), (2) _i The output feature map x of the L-th residual connection can be calculated _L ，

In addition, the convolution layer is used for extracting the characteristics of a local area, different convolution kernels are equivalent to different characteristic extractors, the cavity convolution is adopted, under the condition that the parameter number is not increased, the receptive field of the convolution kernels is multiplied, more data is captured, and the problem that long-time data is difficult to learn can be effectively solved. In order to preserve the integrity of the data as much as possible, the hole convolution does not use the pooling effect and enlarges the receptive field by 0 filling. And residual error connection is introduced on the basis of the cavity convolution, so that the problem of gradient disappearance can be solved.

2) Bidirectional long-short term memory network module

The Long Short-Term Memory Network (LSTM) performs read, write and save operations on input data through a special gating mechanism composed of an input gate, a forgetting gate, an output gate and the like. The gating mechanism can also enable the self-circulation weight value to be changed continuously, and when the network parameters are fixed, the network scale size at different moments can be changed, so that the problem of gradient explosion or gradient disappearance of the circulation neural network can be effectively solved. In order to improve the capability of LSTM to extract features, sequence data are transmitted from two directions by using a hidden layer formed by two independent LSTM units to process forward and backward time sequences, and the sequences output from a forward stream and a backward stream are all connected together to form a bidirectional long and short term memory network (BiLSTM) which can accurately extract time sequence features and solve the problems of gradient loss and difficulty in learning of long sequences.

3) Attention mechanism module

The attention mechanism selectively focuses on important characteristics in a large number of features by means of probability distribution, and ignores most invalid characteristics. The focusing process is reflected in the calculation of the weight coefficient, the weight coefficient of the input characteristic is obtained by calculating the similarity of the output characteristic and the input characteristic, and then the final weight value is obtained by weighting and summing. Therefore, the attention mechanism is added after the serial features, some key input features can be actively collected, and therefore the efficiency of the model is improved. The method comprises the following specific steps:

a)x＝[x ₁ ,…,x _n ]representing N input features, each input feature is scored as formula (4):

s(x _i ,q)＝v ^T tanh(Wx _i +U _q ) (4)

wherein; v. of ^T W is a trainable parameter matrix, U _q As a bias matrix, s (x) _i Q) is a feature x _i Is scored.

The more relevant the input features and output values, the higher their score.

b) The method comprises the following steps Attention distribution value a _i Calculation is performed by activating the function sigmoid:

a _i ＝sigmoid(s(x _i ,q)) (5)

c) The method comprises the following steps The input characteristic X is coded by adopting a soft characteristic selection mechanism, and the coding formula is as follows:

step3: and training and testing the model by using the training set and the testing set. The decomposition model achieves the best decomposition effect by continuously adjusting the model parameters and the network structure. And inputting the actually measured original data into the optimized decomposition model for load decomposition to obtain the decomposition result of each target electrical appliance. As shown in the detailed process diagram of graph load decomposition, power data are measured by a data measurement device and processed to obtain a training data set and a test set. And training the decomposition model in Step2 by using the training set, improving the decomposition effect of the decomposition model by continuously optimizing model parameters and a network structure, and testing the decomposition effect of the model by using the test set. And comparing the test results of each time, and acquiring the optimal model parameters for the decomposition of the measured data to obtain the final decomposition effect graph and index parameters of each electric appliance.

Step4: and analyzing the decomposition effect graph and the parameter index. And analyzing the superiority and inferiority of the decomposition effect through a comparison experiment. And comparing the differences of the decomposition effects of different electrical appliances and factors which may influence the decomposition effects. Further, the performance of the load decomposition model is analyzed by various evaluation indexes. Such as mean absolute error, mean square error, integrated absolute error, precision rate, F1 value, etc.

Claims (2)

1. A non-intrusive load decomposition method of an intelligent building based on a parallel connection network is characterized by comprising the following steps:

firstly, measuring total power data of a certain area and power data of partial high-power controllable load electric appliances by a non-invasive decomposition measuring device, and carrying out data preprocessing to generate a training data set; then, a parallel connection network model is constructed, features are extracted through a parallel connection network combining a cavity residual convolution neural network and a bidirectional long-short term memory network, the characterization capability of the features is improved, an attention mechanism is introduced, redundant information is eliminated, important information is further extracted, and the decomposition effect of the model is improved; then, training a decomposition model by using training data to obtain an optimized decomposition model; finally, inputting the actually measured original data into the optimized decomposition model for load decomposition to obtain a decomposition result, and analyzing the experiment result; the method comprises the following specific steps:

step1: selecting a part of high-power controllable loads of intelligent building subscribers as research objects, collecting total power data and power data of each target electrical appliance by using a non-invasive load decomposition measuring device, and preprocessing the data; further processing is performed using the measured data, generating a model training data set, and calculating a mean value of the measured data set at 8:2, dividing the data set into a training set and a testing set according to the proportion;

step2: constructing a decomposition model based on a parallel connection network;

and step3: and training and testing the model by using the training set and the testing set. The decomposition model achieves the best decomposition effect by continuously adjusting the model parameters and the network structure; inputting the actually measured original data into the optimized decomposition model for load decomposition to obtain the decomposition result of each target electrical appliance;

and 4, step4: and analyzing the decomposition effect graph and the parameter index.

2. The intelligent building non-intrusive load decomposition method based on the parallel connection network as claimed in claim 1, is characterized in that:

in the step2, a decomposition model based on a parallel connection network is constructed;

firstly, after data processing is carried out on input total power, the input total power is respectively input into a cavity residual error convolution network and a bidirectional long-short term memory network for feature extraction; in order to realize multi-angle feature extraction, the traditional layer-by-layer connection mode is abandoned, and the spatial features and the time sequence features of input data are respectively extracted by adopting parallel connection; then, synchronously processing different characteristics to obtain characteristic data with the same dimensionality; then, the feature data with the same dimensionality are connected in series to obtain feature data with composite features; finally, inputting the composite features into an attention mechanism module, further extracting the features, improving the characterization capability of the features, and performing decomposition output through a full connection layer; the decomposition model can be divided into a cavity residual convolution module, a bidirectional long-short term memory network module and an attention mechanism module, and has the following specific structure:

1) Hollow residual convolution module

A Convolutional Neural Network (CNN) extracts features of input data through a plurality of filters; the filters perform convolution and pooling on input data layer by layer, and extract topological structure characteristics contained in the input data layer by layer; although the feature extraction effect of the convolutional neural network is very strong, with the increase of the number of network layers, the problems of gradient disappearance, overfitting and the like can occur; in order to make up for the defects of insufficient load characteristic utilization rate, data loss, gradient disappearance and the like, residual connection is introduced, the idea of cross-layer connection is used for reference, the original input feature mapping and the output feature mapping of the rear layer are added, the feature fusion is completed by utilizing an activation function, and the residual connection enables the front and rear information transfer of the network to be smoother; defining the bottom layer mapping as H (x), adding the input x of the residual error unit, and fitting a residual error function F (x) = H (x) + x through the stacked layers to obtain output; the residual error connection mainly comprises two weight layers, and each layer contains weight parameters; x represents the input and the residual join expression is as follows:

F(x)＝ω ₂ σ(ω ₁ x) (1)

where σ denotes the nonlinear activation function of ReLU, ω ₁ ,ω ₂ Is a weight parameter; finally, adding the input end x and the output result of the second weight layer in a cross-layer connection transfer mode to obtain the final output of the whole residual error structure; the calculation formula is as follows:

y＝F(x,{ω _i })+x (2)

wherein, F (x, { ω [. Omega. ]) _i }) is the residual mapping function to be trained and learned in residual concatenation, { omega } _i Is a set of weight parameters; input feature map x connected by the 1 st residual in the residual structure according to equations (1) and (2) _i The output feature map x of the L-th residual connection can be calculated _L ，

In addition, the features of a local area are extracted by utilizing the convolution layer, different convolution kernels are equivalent to different feature extractors, the receptive field of the convolution kernels is multiplied by adopting cavity convolution under the condition that the number of parameters is not increased, more data is captured, and the problem that long-time data is difficult to learn can be effectively solved; in order to preserve the integrity of the data as much as possible, the hole convolution does not use the pooling effect and enlarges the receptive field by 0 filling. Residual error connection is introduced on the basis of the cavity convolution, so that the problem of gradient disappearance can be solved;

2) Bidirectional long-short term memory network module

A Long Short-Term Memory Network (LSTM) performs reading, writing and storing operations on input data through a special gating mechanism formed by an input gate, a forgetting gate, an output gate and the like; the gating mechanism can also enable the self-circulation weight value to be changed continuously, and when the network parameters are fixed, the network scale size at different moments can be changed, so that the problem of gradient explosion or gradient disappearance of the circulation neural network can be effectively solved; in order to improve the capability of LSTM to extract features, sequence data are transmitted from two directions by using a hidden layer formed by two independent LSTM units to process forward and backward time sequences, and the sequences output from a forward flow and a backward flow are all connected together to form a bidirectional long-short term memory network (BiLSTM) which can accurately extract time sequence features, so that the problems of gradient loss and difficulty in learning of long sequences are solved;

3) Attention mechanism module

The attention mechanism selectively focuses on important characteristics in a large number of characteristics by means of probability distribution, and ignores most ineffective characteristics; the focusing process is reflected in the calculation of the weight coefficient, the weight coefficient of the input characteristic is obtained by calculating the similarity of the output characteristic and the input characteristic, and then the final weight value is obtained by weighting and summing; therefore, an attention mechanism is added after the serial features, some key input features can be actively collected, and therefore the efficiency of the model is improved; the method comprises the following specific steps:

a)x＝[x ₁ ,…,x _n ]representing N input features, each input feature is scored as formula (4):

s(x _i ,q)＝v ^T tanh(Wx _i +U _q ) (4)

wherein v is ^T W is a trainable parameter matrix, U _q As a bias matrix, s (x) _i Q) is a feature x _i Score of (a);

the more relevant the input features and output values, the higher their score;

b) The method comprises the following steps Attention distribution value a _i Calculation is performed by activating the function sigmoid:

a _i ＝sigmoid(s(x _i ,q)) (5)

c) The method comprises the following steps The input characteristic X is coded by adopting a soft characteristic selection mechanism, and the coding formula is as follows:

Images